Anima: What a Company Run by AI Actually Looks Like

Anima is an answer to a simple, slightly uncomfortable question: how much of a small company could actually be run by AI? Not assisted by AI, run by it. Anima is an open-source operating system that puts five specialized agents in charge of the day-to-day work, Growth, Strategy, Operations, Finance, and Quality, and leaves the human one job: review and sign off, in about four to six hours a week. The tagline says it plainly: you sign, AI runs everything.

The idea

Most "AI for business" tools are copilots. They sit beside a person and make that person faster. Anima starts from the opposite end. It assumes the operational execution, the posting, the invoicing, the job assignment, the quality checks, can be delegated to agents, and that the scarce resource worth protecting is human judgment, not human hours. So the human is pulled out of the loop for routine execution and put back in only where it matters: exceptions, strategy, and anything the agents are not allowed to decide alone.

Business logic as configuration, not code

The heart of Anima is what it calls the World Model: a configuration layer that holds everything specific to your company, keywords, vetting rules, budgets, content strategy, as data rather than code. The agents read from that model and act on it. This is the design decision that makes the whole thing work: the business logic is declarative. You change how the company behaves by changing configuration, not by rewriting an agent. And because the World Model is just data, the entire company can be serialized to a portable JSON file and snapshotted, migrated, or seeded, an unusual property for a running business.

The five agents

Five agents divide the work, each triggered on its own rhythm:

• Strategy sets the goal tree and the weekly brief, and dispatches missions to the others. It runs on a schedule, every thirty minutes by default.
• Growth handles SEO, content, leads, and social media, acting on Strategy's dispatch and its own schedule.
• Operations assigns jobs, manages providers, and scores quality, event-driven.
• Finance runs invoicing, margin tracking, and anomaly flagging, on job completion and weekly.
• Quality monitors thresholds, raises alerts, and watches the health of the other agents, continuously.

Each agent operates autonomously within its parameters and escalates anything beyond them to the human.

The human stays at the boundary

This is the part I care about most, and it is where my background shows. I spent eighteen years in control systems, where the rule is never "let it run unattended," it is "give it autonomy, and put the human exactly where authority and accountability have to live." Anima is built the same way. The operator is a Human Proxy with a deliberate time budget of four to six hours a week: a weekly strategy brief, a batch of escalations, a financial summary, a stakeholder call. Everything else runs on its own.

The agents are given real autonomy; the human owns the decisions that carry consequence and signs the things that must be signed. That boundary is not a limitation bolted on for comfort. It is the reason a system like this could be trusted with a real business at all. The same instinct runs through how I think about the agent loop itself: give the loop autonomy, and put the human exactly where the accountability lives.

Under the hood

For the technically curious: Anima is a TypeScript monorepo. A Fastify API and BullMQ workers on Redis coordinate the work; MongoDB holds the World Model and thirteen collections of state, with majority write concern for consistency. The agents run as MCP servers, over stdio locally and HTTP in production, and reason with Claude. The frontend is Next.js. It is built to be operated, not just demoed. The building blocks are the ones I wrote about in the five pillars of agentic AI: instruction layers, MCP for tool access, and agent-to-agent coordination, assembled against a real problem.

Why build it

Anima is a concept made runnable, not a product pitch. The interesting question it explores is not "can AI do this task" but "how much of an organization can you hand over before the human's role becomes purely judgment and accountability." My answer, so far, is: more than most people expect, as long as you are disciplined about where the human sits. The code is open source under an MIT license. If you want to see what a company run by agents actually looks like in practice, the repository is linked at the top of this article.