PolyLabs-research/the_organisation

One prompt in. A company of AI agents out.
the_organisation is a self-hosted, open-source multi-agent AI platform for running a business with autonomous AI agents. Describe an idea and it establishes a working organisation: a CEO that answers to you, an executive team in which every lead runs a standard operating procedure, and agents under each lead doing the work in parallel. They plan, disagree, take it to the board, build, check their own work and report. Then the CEO asks "What would the shareholders want now?" and the next cycle begins. It keeps going until you pause it or the budget runs out.
You are the shareholders. You steer; the CEO enforces it. Spending money and signing go through one seat, the Authorised Signatory, which is you unless you decide otherwise. It is human-in-the-loop by design: agents act inside the privileges you grant, under a hard budget, with every decision on the record.
It runs on Claude, GPT, Gemini, DeepSeek, Qwen, Kimi and anything on OpenRouter, on local LLMs through Ollama, llama.cpp, LM Studio or vLLM, on your Claude Code login with no API key, or in a free simulation with no model at all.
Only Paperclips in simulation: its operating digest, shared commitments and deliveries awaiting verification.
## Features
It is an organisation, not a swarm. A CEO, up to eight executive leads from a library of fifteen (CTO, CPO, CMO, CFO, COO, General Counsel and the rest, or seats you define yourself), and one to eight agents under each. Every lead works to an operating profile, such as Spec → Build → Test → Release, and a phase has to pass the lead's gate before the next begins. Work that fails the gate is sent back.
It remembers the business. A shared calendar and persistent commitments survive cycles. Every job has an owner, next action, dependencies and optional dates. Customer records, a scorecard, financial records, organisational knowledge and incoming events are shared with the agents. Deliveries need evidence and verification.
It can wait. Scheduled reviews, incoming events and calendar dates drive work. Between reviews it sleeps without model calls. The interface and establish command default to hourly reviews; continuous operation remains available.
It governs itself. A lead decides whatever is inside its remit. When two leads' priorities collide, each states a position and the CEO rules, or an independent judge if you seat one. Every decision is on the record, with its reasons.
It does not stop, and it cannot run away. One cycle's reports set off the next round of work, for as long as you let it. A hard budget pauses it. So does a cycle limit. You can pause, stop, steer or change any setting while it runs, and all of its state is plain JSON on disk, so it carries on where it left off after a restart.
A person holds the purse. Agents can draft a contract and choose a supplier, but paying and signing are for the Authorised Signatory, and a person does those, outside the program. The organisation asks, gets on with other work, and treats nothing as paid until you say it is.
Each role gets only what you give it. A Privileges panel sets what each role and its agents may do: search, change records, ask the signatory, read mail, draft mail, sign in, act on accounts, create accounts. Each switch is enforced in code where the thing is done, not by asking the model, and acting on an outside service is off until you turn it on for a role.