We mirror a company into a living twin, rehearse decisions on it, and hand the routine to an AI agent that runs on a heartbeat — every second, minute, hour and day.
Inbox, follow-ups, reports, scheduling, small fixes. A 5–50 person company can't hire for all of it, so most of it just doesn't happen.
Chat windows and copilots wait for a human to prompt every step. They assist. They don't run anything.
Autonomous agents loop, hallucinate and act outside their remit. No audit trail, no limits, no memory of why — so nobody hands them real work.
Messages, tasks, calls, payments and decisions become signed events in one append-only log. Replay it and you get the company as a living model.
Changes are tried against the twin and its real history. Where the agent isn't sure, it escalates to a person with the full story.
An agent that lives by ticks takes the routine, with real tools. Every action feeds back into the twin, so it gets sharper over time.
| 1 s | Reflexes telemetry, alerts, bot commands — no LLM, zero cost |
| 1 min | Triage inbox, routing, quick fixes |
| 1 h | Execution tasks end-to-end with real tools + hourly digest |
| 1 day | Planning larger work, daily manifest, nightly self-training |
Busy levels skip a beat instead of piling up. Only digests travel upward.
ed25519-signed, hash-chained event log. 3.1M events verified end-to-end.
Rights come from verified identity, never from message text. A guard checks every tool call before it runs.
Rules and fast decision models filter before an LLM wakes. Reflexes and quiet beats cost nothing.
Small student models train nightly on real outcomes — and only replace the rules after beating a baseline on fresh data.
1 Owner — full tools
2 Board — work, no secrets
3 Members & agents — scoped
4 Everyone else — reply only
1 s reflex
1 min triage
1 h execution
1 day planning
D0 just this message
D1 the conversation
D2 relevant past episodes
D3 the archive
Memory is private by circle: a board-level conversation can never surface the owner's private chats. The same code serves a two-person studio and a 500-person company — only the table changes.
Every lab is a real business — a founder plus AI agents on shared infrastructure inside Cooperative 42. Each one is a live test bench for the autopilot.
Source: production event store and cooperative portal, 18 Oct 2026. Every number above is a query, not an estimate.
Every decision is logged with what happened next — a reaction, a retry, a fix.
Small models train on that experience and stay in shadow until they beat a baseline.
Repeated failures turn into a proposed rule, sent to the owner for a yes.
Code changes close only with a commit and a passing test. No silent self-edits.
Example from this week: the agent improved its own Telegram behaviour, then tried to close the task without a test. The new gate now refuses that — the change shipped with a commit and a passing test.
The same system our labs run in production. Deployment in the month you join; LLM access through the cooperative's router included. One price for everyone — 10 presale seats.
The autopilot, shared infrastructure, a site, US payments and a founder community. The lab shares a part of its revenue with the cooperative — withheld automatically at settlement.
labs → real experience → sharper autopilot → more customers → more labs
Agents with real tools now finish multi-step work end-to-end. Orchestration, not intelligence, is the bottleneck.
Sub-second decision models and local embeddings let the autopilot think constantly and call a large model only when it matters.
Every company wants autonomy; almost none can audit it. A signed, bounded, self-correcting autopilot is the missing layer.
10 companies on the autopilot. Hardened onboarding, weekly autopilot reports.
Deploy in a day from the portal. Students graduate from shadow to on.
Week and month rhythms: the autopilot plans quarters, not just days.
New labs every month on shared infra; the autopilot as the default operator.
Founder & managing member. Architect of the Simbiot autopilot; chairman of Cooperative 42. US entity, payments and production infrastructure.
Board. 14 years building products and businesses; finance, tax, legal and M&A. Runs Vybirai and Monetix.
Board. End-to-end video production with his own studio; runs ReelsUp, Avatar and xSaleBot.
Sixteen agents in production — sales, content, research, finance, operations — each with its own server, memory and signed track record.
Angel pre-seed · target $500K–1M · 5–8 checks · rolling close
Productize Simbiot for outside companies: self-serve deploy, onboarding, reliability.
More founders through the lab pipeline — each one a new test bench and revenue line.
Fill the presale cohort and turn lab results into a repeatable sales motion.