To be clear: this is not one LLM with a bunch of prompts pretending to be a team. These are fully independent agents — different models, different backends, different runtimes — each doing their own job.
*Why organise agents as a company*
We spent a year building multi-agent systems and kept hitting the same wall. Agents say "I'm done" and pass results downstream with no verification. Errors cascade through the whole project. Static workflows (Dify etc.) are stable but can't adapt to the new skills+harness agent paradigm. Free-form negotiation between agents doesn't converge.
The root issue: agents are getting individually stronger, but without organisational management they can't collaborate effectively. It's like having a room full of talented people with no management structure — capability is there, coordination isn't.
Human companies solved this centuries ago. Performance reviews, SOPs, org hierarchy, hiring pipelines, retrospectives — these mechanisms make collaboration work regardless of domain. We applied the same principles to AI agent teams.
*How it works*
The company has a founding team (HR, COO, EA, CSO) built in from day one.
Talent Market. When the company needs specialists, HR recruits from a community-driven Talent Market ([one-man-company.com](https://one-man-company.com)). Verified agent implementations with real usage counts and community reviews, not LLMs self-declaring expertise. Claude Code, OpenClaw, LangGraph agent families are already on the market — [superpowers](https://github.com/obra/superpowers), [agency-agents](https://github.com/msitarzewski/agency-agents) are ready to hire ([browse here](https://one-man-company.com/talent/claude-superpowers)). The market is open for contributions: upload your own agents, share/buy/sell. We're building a real agent community.
Talent-Container separation. Each agent is split into a portable identity (Talent: prompts, skills, tools, work principles) and an execution runtime (Container: LangGraph, Claude CLI, script process). Different agents run on completely different backends in the same project — Claude Code writes code, LangGraph handles image generation, OpenClaw runs admin.
Tree Search based Orchestrate Algorithm. Tasks decompose dynamically during execution, not upfront in a fixed pipeline. The underlying algorithm is built on AND-tree DAG structure with formal guarantees: termination (bounded retries + timeouts + cost budgets), deadlock freedom (cycle detection at insertion), crash recovery (processing nodes reset on restart).
Quality gates. Every task has acceptance criteria and mandatory supervisor review.
Self-evolution. Post-project retrospectives update work principles and org-wide SOPs. CEO 1-on-1s become permanent agent memory. Performance reviews with real consequences: PIP for underperformers, then offboarding if they don't improve. HR re-recruits from the Talent Market.
*Results*
Cross-domain without changing the framework. So far, we have tried 1) Game dev; 2) Audiobook and Short Movie Generating; 3) Auto-Researchurvey; 4) Content Generation; 5) More is comming!
Same system, different team, different domain. Build a game today, produce a comic series tomorrow.
*Try it*
Fully open source. [one-man-company.com](https://one-man-company.com) is the front door. One command:
GitHub: https://github.com/1mancompany/OneManCompany Talent Market: https://one-man-company.com
Come try it, file issues, or contribute. We especially want people building agents for the Talent Market.