an agent fleet you inherit on day one.

the company brain

Every Praxis agent reads from and writes to a per-tenant, typed memory layer. Not chat memory. Not RAG-over-docs. A structural brain: note types are part of the schema, distillation runs as a worker, conflicts surface in the workspace. Week one the agents are generic. Week twelve they know your customers, your decisions, your projects, the idiosyncratic vocabulary your company runs on. Switching off Praxis means losing that accumulated brain — the moat compounds with every operator interaction.

the autonomy queue

Every action an agent wants to take with side effects declares a decision class. The runtime rejects the call if the class is missing. The class maps to auto, approval, or forbidden per tenant. The coding agent runs in an isolated sandbox with no credentials, capped at 5 files / 50 lines / 16KB / 10 minutes / 24 reads server-side — not in a prompt. There is a kill switch on every agent. This is what lets one operator supervise twenty-five agents touching money, comms, and the repo.

the self-improving loop

The product-pm agent reads the last week of agent runs, decisions, and memory notes; clusters repeat failures; RICE-scores them server-side (the LLM picks impact and effort, the server computes the math); and dispatches the top three to the coding agent. The qa-engineer runs the test suite on every PR inside a sandbox, surfaces the top three failures with suspected root cause, and queues bug-file decisions for operator review. Both leave a traceable improvement record so the next run reads its own scar tissue. Agents fix the agents that fix the company.