Coding is getting cheap. Decision-making is still expensive.

Coding agents have collapsed the cost of implementation. Shipping is accelerating, which puts pressure on the part teams have always struggled with most: deciding what to build, why it matters, and staying aligned long enough to execute. That part has stayed stubbornly manual, even as everything around it gets automated.

Async is a product agent that lives in Slack. It learns a company’s product, customers, codebase, and team through the work already happening there. Ask it what to build next, how to prioritize, where a feature is falling short, what customers are asking for, or what changed this week. It answers with context, the tradeoffs, and references to the conversations and decisions that shaped the answer.

Async also keeps work moving. It spots when a thread is stuck, asks the missing question, and pulls the right people back in. It turns decisions into next steps by drafting scopes and specs, tracking ideas and tasks, and pushing execution forward with coding agents. Everyone operates from the same product context, making it shared and searchable.

We’ve been using Async to build Async at Mainframe. Zero to one product building is a thousand tiny calls and constant questioning. Async holds the running context so I’m not forgetful, and it turns what we decide into actual next steps. It feels like a living, breathing version of the product that’s always there.

Coding agents help teams ship faster. Async helps teams ship the right thing. The future belongs to teams that know what to ship.

Learn more about Async at async.app and follow along @async.

Over one year ago, we set out at Mainframe to explore new ways to interact and interface with AI agents. That vision became Cobot, bringing agents in a shared space where you could accomplish tasks together.

As of today, we’re announcing that we’re winding Cobot down to focus on what’s next.

We’ve greatly appreciated all of you who gave Cobot a try during our early access beta. Your feedback and excitement around our vision gave us the encouragement to keep going. As of recently, we felt the growing desire as a team to find a more specific core focus for a problem we can go deeper on, while still delivering a brilliant experience.

We think we’ve found that thing, and unfortunately it requires that we move on from Cobot. Those of us at Mainframe will keep going, and we’ll share more when we’re ready.

To our existing customers:

• Billing from today is now paused. You won’t be charged moving forward.

• Annual subscribers will automatically receive a prorated refund in the next few days with full access until November 30th, 2025.

• Monthly subscribers will have full access until November 30th, 2025.

• After November 30th, 2025 all user data will be securely deleted in line with our privacy policy.

If you have any questions or need help with your account, contact us at help@cobot.co

Thank you for your trust and support,
All of us at Mainframe

The latest update for Cobot on web and iOS (TestFlight beta 0.1 b83) introduces the Home tab with a feed of your tasks and workflows.

As agents evolve inside of Cobot to be more proactive and work in the background, the feed serves as a living history of their outputs.

When you create a new task or run a workflow in Cobot, you’ll see a post on the feed with the latest updates and progress. It makes it easier to see the activity across all of your agents in one place.

Cobot is in early access beta. A subscription is required during the early access beta.

Ready to meet your new teammates? Sign up today.

Join TestFlight beta

The latest update for Cobot on web and iOS (TestFlight beta 0.1 b71) introduces new Workflow triggers that let your team of agents react and respond dynamically to external events like receiving an email or calendar invite, or when a Linear issue is created or a GitHub PR is opened.

Workflows with triggers are easy to set up. In addition to running recurring workflows on a schedule like daily or weekly, you can now choose external events from Mail, Calendar, GitHub, Linear, and more.

Cobot is in early access beta. A subscription is required during the early access beta.

Ready to meet your new teammates? Sign up today.

Join TestFlight beta

The latest beta adds Liquid Glass for devices running iOS 26. Elements like buttons and cards now act more fluidly and adapt to their environment, taking on optical properties of glass. We’ve also simplified the navigation and updated the Library tab icon.

Improvements and bug fixes

• Fixed a bug in settings navigation

• Fixed a bug where user sessions would lose connection after a period of inactivity

Cobot is in early access beta. A subscription is required during the early access beta.

Ready to meet your new teammates? Sign up today.

Join TestFlight beta

You can now create and edit workflows on Cobot for iOS. Workflows let you build structured, recurring tasks where you can command Cobots using natural language in multiple steps to produce repeatable, reliable outputs.

You can add steps that you describe in words for Cobots to perform as part of your workflow. Pick from Cobots like Mail, Calendar, Notion, Linear, GitHub, and more.

Workflows can be scheduled to run on a recurring basis, including daily or weekly, so that Cobot works for you in the background.

Improvements and bug fixes

• The Notion Cobot is now more reliable when updating pages

• You’ll receive push notifications for messages sent in tasks, workflows, and chats

• After installing a Cobot, you’ll receive a message about what it can do

• Resolved an issue preventing Cobots and Workflows from being installed in some scenarios

• Fixed a rare scenario where the app would log you out during launch

Cobot is in early access beta. A subscription is required during the early access beta.

Ready to meet your new teammates? Sign up today.

Join TestFlight beta

Cobot is your AI teammate for getting work done. It automates the tools you already use, like Gmail, Notion, and Google Calendar, so you can save time on follow-ups, managing your calendar, and task management.

From do to done

Cobot is the first to-do app that does it for you. It’s the simplicity of a to-do list combined with the power of AI agents that we call “Cobots.” Connect your email, calendar, Stripe, Notion, GitHub, Linear, Ramp, and more to let Cobots work together to accomplish tasks.

There's a Cobot for that

Explore the Cobot Library where you’ll find 30+ familiar apps and services that you already know and use. You can add a Cobot to your team in a tap, and discover popular workflows that combine multiple Cobots together.

Two Cobots are better than one

Workflows let you create structured, recurring tasks where you can command Cobots using natural language in multiple steps to produce repeatable, reliable outputs. This is useful when you want Cobots to work together in the background on a regular basis and deliver results straight to your inbox.

Autonomous productivity

We’re excited to see how you put Cobot to work. We believe agents are a tremendously powerful new way to use software, and we’ll continue to build the best experiences around them.

Cobot is now available. A subscription is required during the early access beta.

Ready to meet your new teammates? Sign up today.

Get started

We're excited to release Orchestra v0.0.26, which introduces support for the Model Context Protocol (MCP). This update expands the capabilities of Orchestra agents by enabling them to integrate and use any MCP-compatible server or toolkit.

What’s MCP?

MCP is a simple, standardized protocol for building and sharing tools that AI agents and client applications can easily discover and invoke. MCP allows developers to expose executable functionality—such as system operations or API integrations — in a consistent way, making these tools accessible across multiple agent frameworks and client apps.

MCP has gained a lot of traction due to its unified approach and flexibility, leading to a growing ecosystem of powerful tools available for AI agents.

Orchestra now supports MCP tools

With Orchestra v0.0.26, we've introduced support for MCP tools through the MCPOrchestra adapter. This update allows Orchestra agents to connect to any MCP-compatible server and leverage its tools directly within Orchestra workflows.

This means Orchestra agents can now:

• Use any MCP toolkits, such as filesystem management, web automation, Slack integration, etc.

• Dynamically discover and invoke tools provided by MCP servers.

• Easily integrate external APIs and system-level operations, opening up new use cases.

Example use cases

With MCP integration, Orchestra agents can now easily perform tasks like:

• Filesystem Management: Automatically organize files and folders.

• Web Automation: Navigate websites, fetch data, and automate browser tasks using Playwright.

• Slack Integration: Interact with Slack channels, send messages, and manage communications.

• Custom API Integrations: Quickly wrap and use external APIs like GitHub, Notion, or Airtable.

Why MCP?

The integration of MCP into Orchestra means developers no longer need to manually build custom adapters for each new tool or API. Instead, they can leverage a growing ecosystem of MCP-compatible tools, significantly expanding the capabilities of Orchestra agents who can now immediately discover and use these tools once connected to the MCP server.

Demo

Here's a quick look at Orchestra using MCP tools to organize a cluttered desktop automatically:

Getting started

To explore the new MCP integration, check out the Orchestra documentation and examples provided in the Orchestra GitHub repo. You can quickly set up your own MCP servers or connect to existing ones to expand your Orchestra agents' capabilities.

Other updates

Other recent updates we've made to Orchestra include Braintrust integration for evaluating and monitoring your agents, which you can read more about here in our docs, as well as support for newer language models like GPT-4.5 and Claude 3.7 Sonnet.

We can't wait to see what you'll build with Orchestra and MCP!

Today we’re sharing an early preview of Cobot: an entirely new way for teams of people and AI agents to work together in a shared space. Our vision is to make small teams just as capable and powerful as larger ones by dramatically increasing what your team can accomplish.

We believe that the missing ingredients for getting the most out of agents are making them accessible, collaborative, and familiar. With Cobot, you and your team can orchestrate many AI agents right from your computer or phone to help with common tasks. Cobot makes it easy to check in, delegate, give feedback, and track progress so that things get done.

Cobot is agent-native, multiplayer, and keeps people in the loop. For the first time, different pieces of software can communicate with each other and your team in one place to bridge gaps.

The magic of teamwork

Cobots work together alongside your team. They coordinate amongst themselves using our multi-agent framework, Orchestra, to communicate and handle tasks. As they work with you and your team, they pick up context and knowledge, making them more helpful over time.

They can work in the moment when you need them, on a recurring schedule, or be triggered by external events like receiving an email.

Actions speak louder than words

Cobots don't just chat – they get things done. They connect to your actual tools and services, so they can send emails, update calendars, and handle all sorts of digital tasks. With Auto-Do, they can jump into action the moment you create a task. Set them up once, and they'll handle routine work automatically as you collaborate with and oversee them.

There’s a Cobot for that

We're starting with the essentials – Cobots for email, calendar, and web browsing, much like the apps on your phone. And with the Library, you'll be able to explore and add Cobots to your team that work with all your favorite tools like Stripe, GitHub, Notion, Ramp, Perplexity, and more.

Developers will be able to distribute their agents and place them in an environment together with teams and other agents. Each Cobot comes with permissions and supervision, so you control who can do what.

This is an early preview of Cobot. We’re starting to test with small teams as we continue to iterate and refine the experience.

Ready to meet your new teammates? Sign up for early access.

Get early access

So what actually is an “AI agent”?

Definitions vary, but it can be said that an AI Agent is a cognitive entity powered by a reasoning engine like a language model, situated within an information environment, capable of perceiving inputs, maintaining an internal state, and acting upon the environment to achieve its goals without human intervention.

The term “autonomous agents” has roots in the field of Artificial Life and Artificial Intelligence, and was examined thoroughly in Jeffrey S. Rosenschein’s dissertation at Standford University in 1985; Rational Interaction: Cooperation Among Intelligent Agents.

That said, the agency of these cognitive entities can vary. Reasoning engines can exist in simple loops that are deterministically linked together which don't meet all of the required criteria defining an "agent", or in large networks of communicating agents with greater degrees of autonomy.

Ontologies

Below is a brief walkthrough of some of the architectures that can be built with Orchestra, from more deterministic workflows to hierarchical multi-agent systems with agents that have greater degrees of agency.

Task-Tool Pipeline

This is a pipeline where tasks are assigned a language model and attributes, but not actually assigned a separate agent instance or self-determined tool use. Task outputs are then directly fed into functions, the outputs of which can be fed back into subsequent tasks for further application of reasoning and transformation of information. These are useful when you need an LLM-in-the-loop determining aspects of the flow of the application, but the flow itself is deterministic and entirely known beforehand. Less flexible, but highly predictable, it's somewhat agentic but it doesn't quite meeting all the criteria of an agent.

Imagine a document processing system that handles legal contracts. The pipeline starts with an LLM analyzing the contract to extract key search terms. These terms automatically flow into a legal database search, pulling relevant precedents and regulations. The search results are then progressively summarized by the LLM, translated into different languages for international clients, and automatically saved to the client's file system. While the LLM provides the reasoning, it's following a predetermined path. It supports smaller and more cost effective models and is the least likely to go off the rails. That said, it's not quite meeting the definition of an agent yet.

Task-Tool Pipeline with External Personas

This is a pipeline where the tasks are assigned an agent with a language model but not assigned tools. The task is defined and assigned an agent, and personas can be assigned to multiple tasks. Outputs are still directly fed into functions. This is useful when you have multiple tasks that can be assigned the same agent persona, but still want to maintain strict control of the flow of operations.

For example, a content creation pipeline where the same "editor" agent persona reviews multiple different types of content: blog post drafts, social media posts, and email newsletters. The agent maintains consistent voice/style across different tasks, but the workflow is still predetermined.

Chat Task with an Agent

This is a single task that supports conversation with a single agent. The agent can determine tool calls but there is no discrete, pre-determined task flow.

An example could be a research assistant agent that can engage in back-and-forth dialogue about a topic, using tools like web search or document lookup when needed, but without a predetermined sequence. The conversation flow is dynamic based on the user's needs, and functions are called when needed, according to the discretion of the agent. The conversation flows naturally, with the agent deciding which tools to use based on how the discussion evolves.

Agent Workflows

This is a workflow where the tasks are predetermined and assigned an agent with a language model given tools. Functions are called by agents under their own discretion. This is useful for when agents need flexibility of which function to call, but the sequence of tasks is still predetermined and fixed.

An example of a workflow could be a data analysis agent that decides which statistical tools to apply based on the dataset characteristics - it might run tests, then decide between methods, and select appropriate visualizations, making decisions at each step. The flow of tasks is still pre-determined, but the tools used are up to the agent.

Agent with Callable Pipelines

This is an agent that has tools that it can call that themselves are deterministic pipelines, ideal for repetitive actions and processes that need to be initiated at the right time with the right inputs.

For example, imagine an agent with the ability to call deterministic pipelines. Given different inputs it could decide which pipeline to initiate with the necessary inputs. The pipeline becomes callable with inputs just like a simple tool. It could then pass a URL of a research paper which then initiates a pipeline that downloads the content, splits it into chunks, translates each chunk and saves the accumulated translations to file, or it could initiate standardized workflows like "create new project", which would involve creating folders, setting up documentation, and sending notifications, executing a series of predetermined steps, while the agent maintains higher-level oversight, determining which pipeline to call, when, and their inputs.

Orchestras

Orchestras are defined by having multiple layers, or hierarchies, in the group of agents. Unlike most frameworks that just route messages to and from agents, or simply ask other agents questions, in Orchestra agents can actually create workflows dynamically by decomposing a complex task into subtasks, assign agents to subtasks, and controlling data flows between subtasks. It allows an agent role of Conductor to essentially orchestrate its assigned agents dynamically according to its inputs. Some workflows it issues may involve all of its agents, while others may only require a few tasks to be assigned to a smaller group of its agents.

Think of a software development project where a conductor agent acts like a technical project manager. When given a new feature request, they'll break it down – having a System Designer sketch out the architecture, an API Designer draft the interfaces, Developers handle the implementation, QA ensure quality, and Technical Writers document everything. The conductor orchestrates this complex sequence based on the requirements defined in its input, ensuring the flow follows a logical sequence, managing dependencies between teams, and adapting the workflow based on how the project evolves. They actively coordinate and adjust subtasks for worker agents and managing dependencies based on the project's needs.

Network of Agents

Remaining somewhat theoretical, is the multi-hierarchical network-of-agents. Representing a large network of interacting agents that are capable of carrying out tasks and communicating with each other bidirectionally, these multi-layered hierarchies of agents represents entire organizations or aspects of “mind”.

An example of a network-of-agents could be a newsroom where multiple hierarchies of agents work together: editorial teams (writers, editors, fact-checkers), production teams (formatters, image processors), and management teams (coordinators, quality control) all interact in complex patterns to produce publications. Based on a single input, the conductor layers conduct and orchestrate their own team members in a multi-hierarchical system representing entire organizations.

Orchestra

The range of agent designs in Orchestra - from basic task-tool pipelines to complex networks - shows how the framework can support different levels of agent independence and complexity. Simple designs offer predictability and efficiency for straightforward tasks, while advanced setups allow for adaptive problem-solving through coordinated agent teamwork.

Orchestra's way of organizing agents in hierarchies and breaking down tasks provides a foundation for building AI systems that can grow with your needs. Whether you're building simple LLM-powered workflows or managing complex multi-agent teams, the framework helps you balance capability and control, letting you match agent independence to your specific needs.

As multi-agent systems continue to develop, Orchestra's building-block approach makes it well-suited for exploring new ways agents can work together, while staying true to its core goals of reliability and ease of use. Check out the GitHub repo and the documentation to get started.

fullmoon1, an app to chat with private and local large language models like Llama, is now available on the App Store.2

Making Llama More Accessible

In a previous post by the same name, we believe there’s an opportunity to make models like Llama more accessible to everyone. Right now, the ability to interface directly with these models are mostly limited to developers.

For Mainframe, this means taking the small and mighty Llama 3.2 series of edge models and bridging the gap with a simple app experience that’s intuitive and easy. And with something that’s completely free to use and open source.

We’re optimistic that the Llama series will continue to improve and become increasingly capable, so you can expect all future models optimized for edge devices to be adopted by fullmoon.

Available on iOS, macOS, and visionOS

fullmoon is optimized and compatible with all major Apple platforms, including iPhone, iPad, Mac, and Vision Pro.

After testing and improving with over 3,500 beta testers on TestFlight for the last few months, we’re excited to let anyone install and use models like Llama for free.

Install on the App Store

Both people and agents benefit from breaking down complex challenges into composable pieces. This approach allows agents to tackle tough problems by distributing cognitive load across multiple steps and participants.

Reducing Cognitive Load

When faced with a complex task, the cognitive load can be significant, requiring the LLM to juggle multiple variables, contexts, and potential actions simultaneously. By spreading the cognitive load, Orchestra ensures that each step in the task execution process is focused and manageable.

The approach involves decomposing a large task into smaller, sequential steps or instances. Each step is designed to handle a specific aspect of the task, allowing the LLM to concentrate its reasoning and decision-making capabilities on a single responsibility at a time. This reduces the overall cognitive burden and enhances the clarity and coherence of the LLM's responses.

Cognitive load doesn’t just overwhelm models into repetition or hallucination, it also detracts from the quality of an output for LLMs that aren’t totally overwhelmed.

This method allows the LLM to build upon previous results, maintain a clear execution history, and make informed decisions based on well-defined contexts. By implementing a cascade of steps, Orchestra leverages structured thinking patterns akin to human problem-solving strategies. The result is a more reliable and predictable performance, with improved accuracy and reduced risk of errors.

Orchestration: Conduct and Compose

Orchestra's orchestration capabilities are unique in that they are implemented as tools that can be assigned to any agent, allowing for flexible orchestration patterns. The framework is built around the concept of task decomposition and coordination, where agents can break down complex problems into smaller, manageable tasks on the fly. These orchestration capabilities are implemented as tools that can be assigned to any agent.

The Conduct Tool

The Conduct tool provides a mechanism for enabling agents to delegate tasks to other agents through a tool-based architecture. This delegation capability isn't restricted to a central orchestrator; instead, it's implemented as a tool that can be granted to any agent. This allows for hierarchical delegation patterns, where agents can act as both executors, fulfilling tasks themselves with their tools, and conductors, orchestrating other agents to carry out more complex cross-domain tasks. The act of conducting involves breaking down a task into subtask, assigning the right agents for each job, and specifying which subtask outputs are inputs for downstream tasks. This allows the orchestrator to create complex workflows with one tool-call, and receive the outputs. If necessary, it can reissue a flow of tasks after receiving the results in an iterative and recursive process.

The Compose Tool

The Compose tool represents a higher-level planning layer that pairs with the Conduct tool. Rather than immediately jumping into task delegation, the Compose tool enables agents to first create comprehensive plans when dealing with complex objectives. By separating the planning phase from execution, agents can focus on high-level strategy and optimal task organization. The compose tool itself is only necessary when the task at hand involves a level of complexity, and its separation from the conduct tool allows the orchestrator agents to determine themselves if the request necessitates high-level and comprehensive planning. For tasks that are straightforward delegations (one agent doing a task or two), the orchestrator can directly instruct agents to carry out a task. But for tasks that may involve multiple agents in a more complex request, the orchestrator agent can call on the conduct tool to enrich and expand the orchestration itself by composing before conducting.

Modular Orchestration Patterns

The tool-based orchestration system in Orchestra serves as a building block for creating complex agent hierarchies and organizational structures. Organizations can be modeled with department-level orchestrators managing their specific agent teams, while higher-level orchestrators coordinate across departments. This modularity allows for the creation of specialized orchestration patterns tailored to specific domains or organizational needs.

Agent Communication and Coordination

Agents can communicate bi-directionally, share context, and collaboratively work on complex problems. They can spawn sub-tasks, coordinate responses, and maintain both local and shared state. The orchestrator doesn’t have to carry all of the context from one agents output to the next because it can assign the outputs of one agents work directly as an input in a subsequent task, preventing itself from becoming a bottleneck in the data flow between delegated tasks.

This is true orchestration—not just routing messages, but conducting a symphony of agents working in harmony.

Task Execution

Orchestra's approach to LLM task execution is distinct from traditional function-calling methods commonly used in other frameworks, which impose a significant cognitive load on the LLM. Instead, Orchestra employs a phased execution model that clearly separates concerns through its task execution engine.

Tool Loop Phase

During this phase, tool usage instructions and available tools are clearly delineated, and the original task instruction is insulated with distinct delimiters. The LLM is given only two options: make tool calls or exit the loop. As tool results accumulate within the same context, a clear execution history is built.

Final Response Phase

Once tool usage is complete, Orchestra transitions to the final response phase. Tool instructions and delimiters are removed from the context, and the original instruction is presented cleanly alongside the tool results. This allows the LLM to focus solely on generating the final response without the distraction of tool usage instructions.

This separation of phases reduces the cognitive load on the LLM, as it never has to consider tool usage and final response generation simultaneously. The context becomes clearer, with tool results presented cohesively rather than scattered across function call messages. Decision-making is more focused, leading to more reliable tool usage, better-contextualized responses, clearer reasoning patterns, reduced likelihood of hallucination or confusion, and more predictable agent behavior.

Simplified Tool Integration with Docstrings

Orchestra is built around the idea of managing mental load to support more reliable and predictable agent behavior. The task execution model takes advantage of docstrings to construct prompts for tool usage, offering a more natural and simplified method for instructing the LLM on how to call functions as opposed to the more rigid and complex JSON schemas required by traditional function-calling APIs.

Docstrings are inherently written in natural language, making them more readable and easier to understand for both humans and LLMs. By using docstrings, Orchestra conveys the necessary information about a function's purpose, parameters, and expected behavior without the overhead of detailed JSON schemas. This streamlined communication ensures that the LLM can execute functions efficiently, without being bogged down by the intricacies of a rigid schema.

Orchestra's ability to work with standard Python docstrings means that functions from various sources—whether they're LangChain tools, CrewAI components, Composio utilities, or direct SDK functions—can be integrated into Orchestra agents with minimal or no modification. This compatibility-first approach opens up a wide ecosystem of tools and functions that can be immediately utilized within the framework, reducing the development overhead typically associated with tool integration. Instead of writing custom schemas, adapters, or wrappers for each tool, developers can focus on the functionality and logic of their applications.

Tool-Call Summaries and Rationales

One advantage of Orchestra's approach over rigid function-calling APIs is the ability to extend and customize the schema for tool interactions. For example, Orchestra implements a tool_summaries feature that allows the LLM to provide explanatory rationales for its tool usage decisions. When enabled, each tool call includes a clear explanation of why the tool is being called at that moment, along with function name and parameters.

This kind of metacognitive annotation isn't possible with standard function-calling APIs, which are constrained to a fixed schema of name, parameters, and results. By managing its own tool-calling structure, Orchestra can introduce novel fields and features that enhance transparency and user understanding. Users can understand the LLM's reasoning process in real-time, making the agent's decision-making process more interpretable and trustworthy.

Orchestra represents a shift in how AI agents collaborate and solve complex problems. By focusing on composability, task decomposition, and flexible orchestration through assignable tools, it enables the creation of workflows that mirror human problem-solving strategies.

Visit orchestra.org, view the source on GitHub, and follow @orchestraorg.

In 2018, I imagined a world where the line between human intelligence and machine intelligence blurred entirely, and today that world is quickly taking shape. As 2023 draws to a close, this rapidly emerging arena of AI based autonomous agents will clearly be a highlight. But it’s only the very early stages of the concept. The ideas here represent the extension of this current landscape of isolated agents as they coalesce into AI Agent Clouds.

The future that this explores is one in which AI evolves to create autonomous, intelligent agents that interact and collaborate with each other to form complex networks. In the future, people will be at the center of a functional intelligence-interface layer — a cloud of autonomous AI agents, this is the concept of Agent Clouds.

These AI agents orbit the human nucleus, assuming menial tasks. They interact before we meet, take notes and even check back in after we part, and handle exchanges in between. The cloud is an amplifier of human capacity, making every individual a nucleus of their own unique intelligent cosmos.

Imagine each of us at the center of our own sphere of intelligent agents, that each enable us to make our own best decisions for ourselves. These AI agents don’t work for us; they collaborate, negotiate, and facilitate on our behalf, becoming extensions of ourselves. They’re busy planning and negotiating terms before we even walk into a meeting, and they continue to operate, analyze, and adapt long after we’ve moved on to new ventures.

Building on the vision I started imagining five years ago in a project called Richly Imagined Future, which developed into my contributions to the strategic foresight book Dynamic Future-Proofing by Alexander Manu, this framework further develops the nature of the collaboration and synergy between AI agents, data layers, and humans, ultimately depicting more nuanced interactions.

Instead of viewing interactions in the binary terms of machine-to-machine or machine-to-human, this framework focuses on the dynamic exchange of information and tasks within a network of agents, their associated data layers, and the people they support.

This is in part inspired by the recent development in autonomous software systems like AutoGPT, BabyAGI, Dot, (yes even PizzaGPT). Recent developments in local LLM services like Ollama and LM Studio, tools developed by Langchain, and research done by PlasticLabs on VoE.

2023 and the State of AI Agents

Autonomous agents can be thought of simply as applications that have the ability to think, respond, write and execute code, and autonomously execute tasks​.

Given a goal articulated in natural language, it will do its best to achieve it by dissecting it into sub-tasks and exploring the internet and other tools in a recursive loop.

Centralized services like OpenAI are not the only choice for these tech stacks though. Open source language models like Llama 2 and Mistral can operate directly on local hardware through tools like Ollama and LM Studio, while LlamaIndex can pull relevant context from large datasets for effective context window management, all on-device, without relying on third-party private AI services.

Of course a tech company may create a solution or layer of technology that is more effective than any open-source system involving multiple open-source agents working together, but the open-source AI space is rapidly developing and very competitive. In a domain where local operation allows users to run open-source models on their own devices, local inference with open-source models may better support data ownership and individual sovereignty. In either case, 2023 seems to be the year these agents are beginning to walk.

2018 and the Richly Imagined Future

This echoes the vision I outlined in my Richly Imagined Future project in 2018, where the economy undergoes a transformation by the year 2038:

“Artificial intelligence, machine learning, and digital currencies enabled the machine economy to fully take over the role of labor in society”

In hindsight, these recent AI agents align with that idea that “machines were able to transact, and entirely new layers of the economy formed around machine-to-machine connectivity and exchange.”

Five years ago I imagined a future where our “appliances, phone, home, car… are run by self-executing smart contracts enabling them to operate a series of decentralized micro-businesses”, which begged the question:

“What happens when our machinery, our tools, and our products become financially and functionally autonomous?”

The following framework is an exploration into that same question, a redefinition of the interaction between people and autonomous applications. It is meant to inspire and provide structure to a complex and emerging field.

It ultimately revolves around the question: “what futures emerge in a scenario where artificial intelligence enables a fully autonomous and decentralized machine economy?”. It enables futures where labor shifts focus from the completion of manual tasks toward the deeper exploration of purpose, meaning, and creation.

Our Evolving Relationship with Machines

Before diving into the impact this transformation will have, it helps to look to history for some context (not for pattern prediction though lol).

Industrial machinery obsolesces much of the physical labor done at the time of their invention, which deeply impacted industries like farming, manufacturing, and as a result, the economy and our quality of life.

Domestic machinery obsolesces many time-consuming and demanding activities that were done by hand at the time of their development, altering the patterns of the home and society in positive ways.

Deterministic machinery (the “traditional computer”) obsolesces much of the arithmetic labor that was done by hand at the time of their invention; replacing spreadsheets with CSVs, text processors, calculators, etc.

Intelligent machinery will obsolesce menial intellectual tasks, the boundaries and definition of which has yet to fully emerge.

At every shift, our definition of machinery has advanced, and human capability has increased.

Routine Cognitive Tasks

Routine Cognitive Tasks refer to a subset of intellectual activities that, while requiring mental effort, primarily revolve around repetitive, procedural, or routine-based functions. Unlike the higher-order cognitive functions where human ingenuity, creativity, intent, and long-term strategic thinking are critical, and the purely arithmetic tasks that can be handled by traditional software, routine cognitive tasks are characterized by their predictability and repetitiveness, and their contribution to cognitive fatigue.

Consider these routine cognitive tasks:

Information Digestion

The sheer volume of information available today is becoming totally overwhelming. An LLM that can rapidly and effectively digest this vast amount of data into concise, personalized summaries ensures that you can stay well-informed, without the cognitive fatigue.

Budgeting

Managing personal finances is crucial, yet taxing. An LLM can relieve the cognitive load by handling the nuances of income, expenses, savings, and investments in ways the software we’re familiar with today couldn’t even imagine. By analyzing data, running exploratory data analysis, finding insights, making suggestions, filling out forms, checking over work, LLMs can help relieve so much of the complex, rudimentary tasks involved in staying on top of your finances.

Travel Planning

The dynamic nature of variables involved in route planning, like traffic and weather changes, can make this task daunting. LLMs can continuously learn and adapt from a vast array of data points to provide the most efficient and stress-free travel recommendations.

In this chat example, I worked with ChatGPT-4 with Code Interpreter (itself a component/primitive of their coming GPTs or Assistants) to emulate this process. I sent a request to plan an itinerary for me, my preference to avoid rain, and three files: a flight list of CDG airport in Paris, a document with the weather in Paris, and a list of indoor and outdoor places I want to visit.

It then uses Code Interpreter to import and examine the contents of the files…

It analyzes my request against the data in the files to find the driest 3 day period in my window…

And has put together an itinerary for the 3 driest days I can go to Paris that hits all the landmarks. It didn’t pick up on the request for tickets I was hoping it would, but I wasn’t very clear. With another prompt it would dig into those and help select tickets etc. This is just a simple example of the type of things that will soon be handled entirely by autonomous software.

Comparison Shopping

Product listings and prices change all the time, especially in ecommerce. LLMs can scan multiple platforms in real-time to find the best deals, do produce research, or ensure that choices meet needs, without as much hassle in terms of manual research.

Scheduling Meetings

In professional settings, scheduling can become a game of calendar Tetris, especially when coordinating among multiple stakeholders. LLMs can streamline this process by analyzing everyone’s availability and preferences and propose a few times to all parties, alleviating a lot of the usual back-and-forth.

Many routine cognitive tasks are being automated, not to replace our jobs but to enhance us as people. By offloading these routine tasks onto AI, we can focus on our strengths: empathy driven decision-making, creative drive, ethical considerations, intent, will, and long-term strategic decision-making. The automation of routine cognitive tasks represents not an obsolescence of human capability but an extension of it, increasing our own agency and autonomy, both mentally and temporally.

The Past and the Future

As we look back at the impact that machines have had on human life, from basic hand tools of early civilizations to the complex AI-driven solutions of today, we begin to recognize a pattern. Every era, characterized by the tools and technologies that they employed, paved the way for the next, with each phase being an enhancement, not a replacement.

AI’s role is shifting from prompt-based automation to intelligent autonomous agents that can research, analyze, predict, carry out tasks, and advise. As I mentioned in the Richly Imagined Future project, autonomous organizations are no longer a distant dream but a tangible reality.

Atomic AI Agent Framework

This model is intended to be both a conceptual and practical tool, aiding in system design, anticipating agent behavior and associated social and ethical dynamics, as well as offering a vision for future applications of AI.

The structures of our emerging relationship with AI can be imagined as an atomic structure. At the core of the atom is the human, analogous to the nucleus. Surrounding this nucleus, AI agents come to life like electrons, dynamically appearing and vanishing based on needs and tasks. This singular interplay of human, agent, data, and memory then represents the Atom, whereas the ring of agents itself is the agent cloud, analogous to the electron cloud.

When two such Atoms come into contact, a Dyad emerges, characterized by the shared operations and interactions of their respective clouds of agents. As these relationships become more complex, we see the birth of ‘Molecules’ and eventually ‘Organisms’, representing intricate networks of interdependent Atoms.

The atomic AI agent cloud framework visualizes how humans and AI agents may interact in the future. By understanding the base ‘Atom’, we can anticipate behavior in larger groupings like ‘Dyads’ and ‘Molecules’.

The Anatomy of the AI Agent Cloud

The Human Core

At the center of the Atom is the individual person, the primary element. We constantly interact with the external world, absorbing information from various sources like social media, books, conversations, and experiences, and writing data with our actions, decisions, and creations. We don’t need to capture more data than we do today, but we must aim to control the permeability of this layer between human reality and our collected and stored data.

The Data Layer

Surrounding the human core is the data layer, a dynamic layer that is accumulated over time as users provide data or link accounts together. This layer is strictly metaphorical, in the sense that it is not a data lake, it is a theoretical region, one that supports separate threads or compartmentalized data per agent, per use case.

The Context Layer

Around the data layer is the Context Layer, a conceptual layer that shifts in real-time, organizing and presenting data that is immediately relevant to each agent’s specific task. It’s a dynamic vector store where context is continuously retrieved, structured, and managed. This ensures that the agents engage with information that is not only up-to-date but also specifically curated for the task at hand, improving the effectiveness and relevance of the agent’s functions.

The Agent Cloud

Beyond the context layer lies the Agent Cloud, a sophisticated network of AI-powered autonomous agents, each specialized in specific tasks. These agents continuously interact with the data layer, pulling context and chaining prompts to perform specialized functions.

Together, these agents form the agent cloud, a membrane between the human core and the vast amounts of information we deal with daily. They handle a wide range of tasks, making life less demanding and more aligned with our choices and needs.

Core Concepts

Dynamic Data Stream

The surrounding data is a stream that flows, changes, and adapts. It incorporates continuous inputs from the user’s experiences, agent memories, recent interactions, and other contextually relevant data. This stream also ensures synchronization between the human core and the Agent Cloud, providing real-time updates that reflect any new developments in one’s life.

Adaptive Prioritization

The Agent Cloud recognizes and responds to the relevance and urgency of tasks as they arise. Whether it’s an unplanned business trip or an unexpected family event, these agents quickly adjust their priorities.

Cumulative Learning with Memory

Agents compound their knowledge over time, refining strategies based on previous experiences. A pivotal component of this learning is the agent’s memory, which allows them to avoid repeating mistakes and build on successful strategies. By incorporating principles like those from the paper on Violation of Experience by PlasticLabs, agents can better understand the importance of aligning actions with past experiences and user preferences.

Inter-Agent Collaboration

At the heart of the Atom is the seamless collaboration between agents. These agents aren’t isolated entities; they function as part of a collective. They share information, strategies, and insights, ensuring that tasks are handled efficiently. For instance, if one agent identifies a user’s interest in cooking Italian cuisine, another agent responsible for shopping can procure ingredients for an Italian dish. It’s a series of collaborations where agents complement each other’s functions.

Future 1: Atomic Agent Cloud

Avery’s day began by listening to a concise summary of the news, key events, meetings, and tasks for the day with context drawn from her news, and social feeds and notifications, as well as her day’s schedule.

On her way to work, Avery received a message from old friends, inviting her for a spontaneous trip over the weekend. Before she could even start wondering about the logistics, her travel planning agent had already analyzed the options. Considering her preferences, recent conversations, and past travel experiences, the agent presented Avery with a suggested itinerary and even incorporated potential impacts on her finances and upcoming plans.

While at work, Avery faced a deluge of emails from various teams, but her communication agent quickly sifted through them all, summarizing key points into a concise brief, ensuring she didn’t miss any key updates. Later, as she prepped for her team’s weekly meeting, her meeting agent had already organized the context of the last meetings, setting a tentative agenda and even pulling up relevant pre-read materials.

Post work, while contemplating attending an evening workshop, her schedule agent displayed the potential clashes and solutions. It even highlighted a few colleagues who were interested, a nudge from her networking agent.

In the evening, as Avery thought about dinner, her recipe manager agent suggested a few dishes from her favorite recipe book. Cross-referencing with her food inventory manager, it ensured she had most of the ingredients at hand. However, for a particular Italian dish she was leaning towards, her shopping manager highlighted a few ingredients she was short of and offered to place an order or schedule a shopping reminder.

This collaboration between her agents ensured that tasks were simplified, leaving her with just a few streamlined decisions to make. These decisions then activated a series of automated actions, alleviating the mental clutter often associated with daily chores.

The Dyad

A Dyad is a structured, secure, and dynamic relationship between two atoms, designed to facilitate a multitude of tasks and objectives for the individuals they represent.

Anatomy of the Dyad

Unified Interaction Space

This will serve as the primary communication channel for the two agent clouds. All real-time communications, from initial negotiation to ongoing interaction, can happen here. This unified chat approach minimizes overhead and keeps the interaction straightforward.

Shared Context Database (RAG-enabled Data Repository)

This temporary, sandboxed vector database allows agents to pull contextually relevant data directly into their decisions and interactions. For the conversation’s requirements, this data is readily available and indexed for quick retrieval. Any more extensive datasets or resources that are required can also be pulled in and temporarily housed here. This data is then made accessible to the agents for their interactions via retrieval augmented generations.

Interaction Protocol

This is a set of rules and guidelines that dictate how the agents from both clouds can interact, make decisions, and negotiate. These mechanisms enforce consent and privacy, prioritizing human agency by enforcing strict permissions on data sharing and requiring explicit permissions and human decisions / approval for any meaningful actions.

A clear definition of what constitutes a ‘meaningful action’ then becomes necessary, and while I can’t offer one, I can say at least that the decision of clicking a button on a website to read the contents of a page wouldn’t be a ‘meaningful action’, but ordering something on your behalf (like spend access) would certainly be.

Event Log

A chronological and auditable record of interactions, decisions, and changes made during the life of the Dyad.

Core Concepts

Feedback Loop

A mechanism for each Agent Cloud to evaluate the success of the interaction, and to learn from it for future Dyads. I’ve mentioned VoE as a good example of the proven frameworks available for autonomous self-improvement based on interaction feedback.

Memory Weightage

Assigns importance scores to memories based on factors such as emotional significance, frequency of interaction, or potential future benefits. Memories with higher weightage are prioritized and retained, ensuring that the Agent Cloud’s storage remains efficient and relevant.

Future Interaction Potential

Evaluates the likelihood of repeated interactions with an entity. For instance, if the Agent Cloud predicts that you’ll interact with a specific colleague or friend more often, the information from past interactions with that individual might be stored for a more extended period, whereas if it’s a standard, anonymous, and ephemeral interaction then it might assign a lower significance to the future interaction potential.

This helps to avoid the agent memory bloating with every single interaction you’ve had, but also helps it pay more attention to those whom you’ve just met but are likely to meet again, like a first meeting with a new colleague.

Retention Review

Periodic evaluation of stored memories. Memories whose relevance has diminished might be compressed, archived or discarded, ensuring the Agent Cloud’s storage stays light.

Memory Transparency & User Pruning

An essential aspect. Users can manually prioritize certain memories for retention or deletion, ensuring they remain in control of what their Agent Cloud remembers.

Phases: Interactions Between Two Agent Clouds

Initiation & Data Handshake: This phase involves two agent dyads becoming aware of each other’s presence and engaging in an initial data exchange. They share contextual information about their respective users without compromising private data. This sets the stage for the communication between the two humans

Alignment & Objective Setting: The agent dyads collaborate to identify common goals, interests, and objectives of the upcoming interaction. This could be setting the agenda for a business meeting, understanding the learning goals of a student, or gauging the cultural nuances to be addressed in a cross-cultural conversation.

Intelligent Context Layering: Here, the agents assess and superimpose layers of context onto the interaction. The decisions to pull in data should align to user settings, and data layers imported into the temporary interaction window could be as broad as general trends or as specific as the meeting minutes of a specific meeting a few months ago. It’s all about making the interaction richer and more meaningful by adding layers of information that might be of interest, as the conversation unfolds.

Real-time Facilitation: Here, the agent dyads actively assist during the live interaction. This involves tasks like real-time translation, instant recall of past interactions or relevant information.

Feedback & Learning: Post-interaction, individual agents analyze the conversation’s outcomes against their set objectives, hypotheses and expectations. They learn from any discrepancies, updating their mental models for future interactions.

Molecule

If the Dyad represents the future of interpersonal interactions in a digital-first world, then the Molecule extends this concept to group dynamics. The Molecule is a complex, multi-faceted relationship among multiple Agent Clouds, encapsulating small friend groups, teams, small companies, or even nonprofits. In the same way that atoms form molecules in the physical world, individual Agent Clouds connect to form Molecules in the digital realm.

Note: the concepts going forward are less technical and more conceptual as we reach outside of the bounds of what I can imagine

Structural Components of a Molecule

Collective Context Layer

In a Molecule, the first structural element is the collective context layer. This layer aggregates essential, non-intrusive data from each individual Agent Cloud involved. This aggregated data provides a comprehensive snapshot, offering insights that guide the actions of the Molecule.

Group Priority Mapping

After forming a collective data layer, the Molecule undergoes a “Group Priority Mapping.” This step involves analyzing the goals, priorities, and preferences of all the participating Agent Clouds to formulate a collective set of objectives and actionable tasks.

Governance Protocol

Every Molecule has a governance protocol that determines how decisions are made, how tasks are allocated, and how conflicts are resolved. This ensures a balanced and fair interaction among all the constituent Agent Clouds.

Molecular Processes

Each agent scans the collective data layer to extract information relevant to its specific tasks or goals.

Agents maintain data integrity and privacy by only accessing or modifying data for which they have explicit permission.

Agents continuously update the collective data layer with new information, allowing the Molecule to adapt dynamically.

Agents analyze data not just for content but also for context, considering the source, timing, and relational aspects to other data.

Transparent Ledger

An open yet secure log that captures all interactions, transactions, and decisions made within the organism, providing a history and accountability layer.

Tasks Handled by a Molecule

Project Management — Coordinating timelines, milestones, and tasks among multiple participants to achieve a common goal throughout multiple projects.

Delegation & Resource Allocation — Identifying and distributing resources like allocating workloads, expertise, or assets to fulfill group objectives.

Molecule Examples

Family Units

A small family could function as a molecule. Each family member’s Agent Cloud could coordinate to manage household chores, family outings, meal plans, and so forth.

Book Clubs

A book club where everyone reads the same book and meets once a month could also be a Molecule. The agents could decide on a book, set reminders for reading targets, and arrange meetings.

Nonprofit Organizations

A small nonprofit focused on community aid could represent a molecule, where tasks like fundraising, event planning, and volunteer coordination are managed.

Organisms: Inter-Molecule Community Networks

Just as individual Agent Clouds engage in Dyads, Molecules might also interact. The organism could represent a molecule’s interactions with other molecules, negotiating and establishing links, adding another layer of complexity to the system.

Anatomy of an Organism

Clustered Vector Pools

Clustered Vector Pools would be regions within the Organism where similar or related Vectors from various Molecules converge, facilitating easier access and reference for common tasks.

Network Alignment

Network alignment would be necessary to serve as the Organism’s communication backbone, ensuring seamless alignment among all Agent Clouds. They could propagate directives and policies to every Agent Cloud, enable high priority message broadcasts for urgent notifications, and maintain a persistent ping mechanism for updates. They might even need to conduct periodic alignment checks to ensure Agent Clouds are synchronized with the Organism’s latest knowledge and utilize feedback loops to gauge and refine the effectiveness of communicated directives and updates.

Boundary Sensors

The Organism can have specialized boundary sensors at its periphery that detect external stimuli or changes in the environment, triggering responses or adaptations within the Organism.

Ensuring Alignment

The Agent Cloud concept brings with it many questions. It’s not necessarily a proposal, but something of an anticipated future. One of the goals of anticipating downstream futures is to identify risks and potentially detrimental outcomes in order to help guide the design of these systems. Thinking big, and dreaming of optimistic futures is crucial to building a future we want to live in, but we must do that while avoiding poorly designed technological infrastructure.

The augmentation of individual capabilities through AI agents requires a strong and tuned moral compass; when AI magnifies human intent, both benign and malicious, the ethical dimensions multiply.

Drawing from “Dynamic Future-Proofing,” we have to ask, now even more than before: Who holds agency over these agents? What mechanisms deter misuse? How do we manage consent and privacy? How do we keep those protections at the forefront of the tech? This future calls for a definition of the moral architecture that should steer semi-autonomous agents.

User control and alignment with personal values are key considerations. An essential aspect of this control is the ability for users to define and fine-tune how their AI agents operate, ensuring that the dynamic is that of machine assistance driven by human intent and decision making. Violation of Expectation (VoE) and metacognitive prompting, concepts and methodologies illustrated by Leer, Trost, and Voruganti at PlasticLabs, would contribute in this regard. Further granular settings should offer control over data sharing, contextual operations, and task-specific consents.

The dynamic between AI agents and humans should be rooted in continual refinement and feedback. Transparent action logs and feedback loops are necessary to offer users a clear view into their agents’ activities, allowing for adjustments as needed, incorporating the processes and methods of explainable artificial intelligence (XAI) that enable people to understand and trust the operations and outputs of their LLMs and Agents.

The introduction of community-shared ethics and collaborative setting adjustments further paints a picture of a future where the collective wisdom of many aids individual users in sculpting their AI’s ethical and operational landscape.

As AI becomes a mainstream navigator in our lives, there’s a looming shadow — the risks of relying on models that have unintentionally absorbed the biases and inaccuracies of humans as filtered by the information available on the internet, which is itself heavily biased. This gap between an accurate, real world view and a biased one, is a gap that risks widening with further integration of AI systems without first addressing those inaccuracies to avoid leaving some behind in the wake of this rapid progress. Addressing this calls for AI ecosystems that are not just intelligent and capable, but also broadly accessible, ensuring technology uplifts everyone, not just a select few.

Conclusion

In our exploration of human-agent collaborations, from the foundational ‘Atom’ and collaborative ‘Dyad,’ to the complex ‘Organisms’, we’ve recognized the profound symbiosis between humans and machines. The vivid examples, spanning from day-to-day tasks like travel planning to complex endeavors such as business operations and delegation, depict a near future where the boundary between human endeavors and machine intelligence is seamless.

“The machine is financially and functionally autonomous and transactions are made on a H2H, M2H, and M2M basis”: These agents could potentially handle transactions in intra-agent operations between agents within a single agent cloud or inter-agent operations between agents of different clouds.

However, understanding this emerging landscape is just the beginning. It’s imperative to grasp the broader context of these developments. The trajectory we’re on doesn’t merely suggest iterative enhancements but hints at a paradigm shift in human-machine collaboration.

Strategic foresight is not about predicting the future but about preparing for multiple probable futures. Our current endeavors are indicators, signposts pointing towards these futures. By acknowledging and interpreting these signs, we position ourselves to design for these futures, and in doing so, actively shape them. It’s about having the audacity to envision a world where our digital counterparts don’t just complement, but amplify our capabilities.

Our journey through the Atom, Dyad, Molecule, and Organism has provided insights and frameworks. Yet, the real challenge lies ahead: to harness these insights, anticipate challenges and opportunities, and architect a future where our combined potential, both human and machine, is realized to its fullest extent, while keeping people at the center. It’s not just about solving problems, but enhancing human agency while protecting privacy and consent.

Consider this a call to action to anticipate this future, and begin designing and engineering the best path forward.

Appendix of Terms

Retrieval Augmented Generations (RAG): A method that integrates embedding models that transform text into vectors in multi-dimensional point space with knowledge retrieval mechanisms, allowing the model to pull in additional relevant context when generating responses to improve accuracy and relevance.

Vector Database: A database that stores data in the form of vectors, representing points in a multi-dimensional space. Such databases are optimized for high-speed similarity searches, making them useful for tasks like information retrieval or machine learning models.

Relevant Context: Pertinent information retrieved from embedding models during a RAG process, ensuring generated responses are accurate and contextually-aware, enhancing the model’s understanding and relevance to the query.

Intelligent Synchronization: The process by which various agents, systems, or databases update and align their information, ensuring that all entities have access to the most current and relevant data.

Dynamic Delegation: A method where tasks are assigned to agents based on their capabilities, availability, and the current context. This ensures efficient resource utilization and timely completion of tasks.

Agent: A digital entity or software designed to perform specific tasks autonomously.

Agent Cloud: A collection or network of agents that operate in tandem, sharing resources and information. They work collaboratively to handle complex tasks and decisions based on collective intelligence.

Atom: The foundational unit in the AI Agent Cloud concept, composed of the Human Core, Data Layer, and Agent Cloud.

Dyad: A complex unit formed when two Atoms interact, leading to the sharing of agents and data layers.

Molecule: An even more intricate structure that emerges from the interaction of multiple Dyads. This involves several humans, their respective data orbits, and an interconnected cloud of agents.

Dynamic Future-Proofing: A conceptual approach emphasizing the need to anticipate future developments and challenges, ensuring systems and strategies remain relevant and effective over time.

Decentralized Machine Economy: An economic structure wherein machines or digital entities have their own autonomy, driven by algorithms and devoid of centralized control, enabling novel human to human (H2H), machine to human (M2H) and machine to machine (M2M) relationships to emerge. (2018, Richly Imagined Future, P. Pagé).

Collective Data Layer: A shared repository of information accessible by multiple agents or entities within a network, fostering collaboration and informed decision-making.

H2H (Human-to-Human): Traditional interactions that take place directly between two or more humans, without the intervention of automated systems or digital agents.

M2M (Machine-to-Machine): Refers to direct communication between autonomous agents without human intervention. This can involve anything from two computers exchanging data to complex systems where multiple agents collaborate.

M2H (Machine-to-Human): Describes the interaction between machines (typically computational devices or software) and humans. This can be anything from a user interface on a computer to more advanced interactions involving AI or intelligent agents assisting human users.

Agent Cloud Atomic Model: A more advanced and intricate framework that emphasizes the collaboration and synergy between agents, data layers, and humans. Instead of viewing interactions in the binary terms of M2M or M2H, the Agents Model focuses on a dynamic and fluid exchange of information and tasks within a network of agents, their associated data layers, and the humans they help. This model facilitates more nuanced interactions, ensuring effective task delegation, learning, and adaptation based on changing contexts and needs.

2023: State of AI Agent Links — ​1​, ​3​, 4, ​5, ​6, ​7, ​8​

Dynamic Future-Proofing: Pagé, P. (2021) In A. Manu (Ed.), Dynamic Future-Proofing: Integrating Disruption in Everyday Business (pp. 143–151). Emerald Publishing Limited.

https://books.emeraldinsight.com/page/detail/dynamic-future-proofing/?k=9781800435278

https://publish.obsidian.md/philippepage/Work/Agent+Clouds

Today, Mainframe is open-sourcing and releasing an early beta preview of fullmoon, an experimental app to chat with private and local large language models.

With the release of Llama 3.2’s lightweight 1B and 3B models, the ability to run a highly capable model locally with privacy where data never leaves the device is now possible. With fullmoon, we want to bring the best experience for chatting with Llama to everyone, making it more accessible.

Running these models locally comes with two major advantages. First, prompts and responses can feel instantaneous, since processing is done locally. Second, running models locally maintains privacy by not sending data such as messages and calendar information to the cloud, making the overall application more private.

— Llama 3.2: Revolutionizing edge AI and vision with open, customizable models

fullmoon

fullmoon1 is an experimental iOS app to chat with Llama that’s optimized for Apple silicon and works on iPhone, iPad, and Mac. As of today, you can install and chat with Llama 3.2 1B and 3B models. Your chat history is saved locally, and you can customize the appearance of the app.

It’s available now on TestFlight as an early beta preview. Your feedback is welcomed and appreciated.

Open source

We’ve open-sourced the fullmoon iOS app on GitHub, built with SwiftUI. fullmoon is made possible by MLX Swift from Apple. MLX is an array framework for machine learning research on Apple silicon, brought to you by Apple machine learning research.

We’d love your help, support, and contributions in continuing to evolve fullmoon and making all current and future large language models that are optimized for on-device even more accessible.

You can find us @mainframe and mainfra.me. If you’d like to join us, email careers@mainfra.me.

What’s a computer?

We’re at a point where we can question all prior assumptions made about computers. How should they work and how will we interact with them?

Mainframe wants to redefine human-computer interaction with a future operating system where agents are the new apps.

By 2030, we believe that:

• AI is the default

• Software generates itself

• Agents are the new apps

Large language models have transformed the perceived capabilities of computers. They can now think and reason alongside us. What does it look and feel like to operate in a new environment where the bicycle for the mind has an engine?

Software is changing and anyone can make anything. Mainframe’s mission is to change the world's relationship with computers.

You can find us @mainframe and mainfra.me. If you’d like to join us, email careers@mainfra.me.