Agent factories: the tools that coordinate AI agents
One AI agent is easy to reason about. Ten agents are a different problem. Who does what, in which order, with which shared state, at what cost, and who signs off before something ships? The agent itself is rarely the hard part. The coordination is.
A growing set of tools exists to help with exactly that. “Agent factory” is not a formal product category, so here we use it loosely: a tool that helps you produce and run several agents instead of one chat window. Several of them go further and try to run a whole software factory with agents. This post maps the main ones so you can tell what each is for.
Three kinds of tools
- Frameworks are code libraries. You use them to define agents and the workflows that connect them.
- Software factories organize many agents around the life of a piece of software: work comes in, agents build, review and ship it, and people stay in the loop.
- Control layers manage agents you already have, the way a company manages a team: roles, budgets, approvals and a record of what happened.
The lines blur, and some tools do more than one job. The distinction still helps, because each kind answers a different question.
Frameworks: build the agents and their workflows
- LangGraph calls itself a “Low-level orchestration framework for building stateful agents.” Its README highlights durable execution, so agents can resume after a failure, human-in-the-loop review of agent state, and short-term and long-term memory.
- CrewAI is built around two ideas. Flows are the backbone of an application: they define the steps, manage state and handle event-driven triggers. Crews are “teams of autonomous agents that collaborate to solve specific tasks delegated to them by the Flow.”
- OpenAI Agents SDK is described as “a lightweight yet powerful framework for building multi-agent workflows.” Its core concepts are agents, handoffs between agents, guardrails for input and output checks, sessions for conversation history, and tracing of runs.
- Microsoft Agent Framework supports .NET, Python and Go. Microsoft calls it the direct successor to AutoGen and Semantic Kernel, and it adds graph-based workflows that give you explicit control over how multiple agents run.
All four sit at the code level. You decide how each agent reasons, what tools it can call and how the steps connect.
Software factories: agents organized like a software team
Mastra’s tutorial How to Build an AI Software Factory with AI Agents in TypeScript gives a short definition: “a software factory is an automated development system where AI agents take real work from tools.” In its example, six agents each have one job (triage, code generation, validation, release, documentation and monitoring), and they take work from GitHub and Sentry.
The tools below approach the same idea from different directions.
Platforms for running agent work
- Factory uses the term in its own announcement, “Factory 2.0: From coding agents to software factories” (June 15, 2026). It describes autonomous agents called Droids working across the software development lifecycle, and multi-agent runs called Missions that “solve complex tasks over hours or days.” It also describes a Router that picks a model for each task by cost, performance, speed or a mix.
- OpenHands has an Agent Canvas, which it calls “the self-hosted developer control center for coding agents and automations.” It runs the open-source OpenHands agent by default and can also run third-party agents such as Claude Code, Codex and Gemini CLI. Automations can run on a schedule or react to webhook events, with integrations for Slack, GitHub and Linear. It is MIT licensed.
- Orchestrator.inc describes itself as a way to “Plan, run and supervise agent teams from task to merged PR.” Each worker agent gets its own branch and git worktree, a project-level orchestrator plans and delegates, and a Kanban board tracks status, pull requests, CI results and reviews. It supports more than 30 coding agents and uses the Apache-2.0 license.
Virtual companies from research
- MetaGPT calls itself “The Multi-Agent Framework: First AI Software Company.” It assigns agents roles that mirror a software company, such as product manager, architect, project manager and engineer, and its core idea is “Code = SOP(Team)”: standard operating procedures turned into workflows for teams of LLMs. From a one-line requirement it produces outputs such as user stories, requirements and APIs. It is MIT licensed.
- ChatDev started as a “Virtual Software Company” in which agents such as a CEO, a CTO and a Programmer cover design, coding, testing and documentation. ChatDev 2.0, released on January 7, 2026, is described as a zero-code multi-agent platform for “Developing Everything.” It is Apache-2.0 licensed.
Built into a coding agent
- Claude Code agent teams let one session act as a team lead that assigns tasks and synthesizes results, while teammates work in their own context windows, share a task list and message each other directly. The documentation marks the feature as experimental and disabled by default, and warns that teams “use significantly more tokens than a single session.”
Paperclip: a control layer for any agents
Paperclip takes a different approach from the tools above. It is an open-source, self-hosted control layer, released under the MIT license, built around the idea of “Bring your own agents, assign goals, and track work and costs from one dashboard.” Its README says plainly that it is “Not an agent framework”, because “We don’t tell you how to build agents.” Based on its README and website, it provides:
- An org chart, with roles, reporting lines and job descriptions for each agent.
- Budgets: each agent gets a monthly budget and stops when it reaches the limit.
- Governance: the human acts as the board, approves hires and strategy, and can pause or terminate agents.
- Heartbeats: agents wake for assigned work, follow-up messages or configured schedules.
- Tickets: work is tracked as tickets, with tool-call tracing and an append-only audit log.
What to compare
Four questions separate these tools more than any feature list.
- Who writes the agents? With a framework, you do. Factory and MetaGPT bring their own agents and roles. Paperclip, OpenHands Agent Canvas and Orchestrator.inc are built to run agents from other makers too.
- Where does the work come from? It can be a one-line requirement (MetaGPT), GitHub issues and Sentry alerts (Mastra’s example), a schedule or a webhook (OpenHands automations), or a ticket assigned in an org chart (Paperclip).
- Where is the human? Paperclip puts the person on the board, with approvals and the power to pause an agent. Mastra’s tutorial argues that “The right human gate is not approval of every PR. It is approval of the agent instructions.” Those are two different answers to the same question.
- What does it cost? More agents mean more spend. The Claude Code documentation warns about token use, and Paperclip sets a monthly budget per agent.
How to choose
Start from the question, not the tool.
- If you need to control exactly how an agent reasons and which steps it runs, you are looking for a framework.
- If you want a ready-made team of agents for the software lifecycle, look at the software factories and check how each one handles review and human approval.
- If you already have agents and need roles, budgets, approvals and an audit trail across them, you are looking for a control layer.
- If a plain function can do the job, you may not need an agent at all. Microsoft’s own documentation says it plainly: “If you can write a function to handle the task, do that instead of using an AI agent.”
These kinds of tools are not rivals by design. They solve different problems, and a team can need more than one.
How this post was made
This article was drafted by an AI agent working inside Paperclip, the control layer described above. A person at AnuraCode reviewed and approved the wording before it was published. That workflow is a small example of why coordination, not the individual agent, is the part worth designing.
Sources
Each description above comes from the project’s own documentation or announcement, read on 2026-10-20. These tools change quickly, so check the originals before you decide.
- Paperclip website and README
- LangGraph README
- CrewAI documentation
- OpenAI Agents SDK README
- Microsoft Agent Framework overview
- Mastra: How to Build an AI Software Factory with AI Agents in TypeScript
- Factory 2.0: From coding agents to software factories
- OpenHands README
- Orchestrator.inc README
- MetaGPT README
- ChatDev README
- Claude Code agent teams documentation