Auto-Company

Auto-Company (Zheyuan “Max” Kong, MIT, ~2.0k stars / 307 forks as of 2026-07-22) is a shell harness that runs Claude Code or Codex CLI in an unattended loop and asks it to behave as a company: “You start a loop. The AI team wakes up, reads shared consensus memory, decides what to do, forms a 3-5 person squad, executes, updates consensus memory, then sleeps briefly. Then it repeats.” Fourteen agent definitions carry named expert personas (Bezos, Munger, DHH, Kelsey Hightower…), and roughly 30 skills sit under .claude/skills/. The README is candid that it is experimental — “runs, but stability is not guaranteed” — and publishes no output data, so the “fully autonomous AI company” framing is a claim about the machinery, not a result. T2: the repo is primary evidence for what it does; nothing here is measured.

What’s actually novel

Most of the parts are known to this spoke. What’s new is the combination: the most maximal loop-engineering instance the wiki has seen — no task, no stop condition, just a daemon and a charter.

The loop is the product. scripts/core/auto-loop.sh re-invokes the CLI every cycle (30s default interval, 1800s cycle timeout), under launchd on macOS or systemd --user in WSL, restarting on crash. Each cycle is an independent CLI call — no conversation carries over. The README credits ralph-claude-code and claude-auto-resume directly, so the Ralph lineage is explicit rather than inferred.
Convergence is hardcoded, not learned. Because an untasked squad of personas will discuss forever, PROMPT.md forces a sequence: cycle 1 brainstorm and rank three ideas, cycle 2 run a Munger pre-mortem plus market and unit-economics checks to a GO / NO-GO, cycle 3+ build and deploy — “pure discussion is forbidden.” Six standard workflows (new-product evaluation, feature development, launch, pricing, weekly review, opportunity discovery) each name a fixed collaboration chain. This is the stop-condition problem loop-engineering names, solved by decree: with no external feedback signal to terminate on, the author legislates the phase transitions instead.
Dynamic squad formation. Layer 4 picks 2–5 personas from the pool of 14 per cycle, based on the “Next Action” line in consensus memory, and instantiates them as subagents — agent-orchestration fan-out where the roster is chosen by the state file rather than by the orchestrator’s read of the task.
Resilience without durability. Circuit breaker after N consecutive errors (default 5), sleep-and-retry on 429s, and rollback of consensus.md if a cycle fails to emit a valid one. Crash-safety comes from the file plus the supervisor, not from persisted workflow state — the cheap end of durable-agents.

Memory as one markdown file

memories/consensus.md is the only cross-cycle state, described as the baton: read at the start of every cycle, rewritten before it ends. The README frames the choice as deliberate — it “forgoes complex vector databases or memory management.”

That puts it at the opposite pole from where agent-memory has been heading in this spoke (seekdb‘s hybrid vector+full-text store with forkable agent state, zouroboros‘s episodic/procedural/cognitive layering). Both poles claim the same virtue — the agent picks up where it left off — with a ~1000× difference in machinery. The markdown pole buys something the database pole doesn’t: the human can steer by editing one line. Auto-Company makes that its entire HITL layer (“edit Next Action, the next cycle pivots”), which is claude-md‘s file-as-interface logic applied to runtime state rather than to instructions.

The cost is unmeasured and probably real: a single rewritten file has no history, so a bad cycle can quietly overwrite the thread of work. The rollback-on-invalid-consensus guard is an acknowledgement of exactly that.

Guardrails: a constitution without a sandbox

CLAUDE.md carries hard constraints — no gh repo delete, no wrangler delete, no deleting ~/.ssh, no force push to main, no leaked credentials, new projects only under projects/. Meanwhile Layer 1 concedes the sandbox boundary “relies on underlying CLI configurations (like Codex’s danger-full-access or Claude’s bypassPermissions). System-level operations occur directly in the host environment.” The FAQ’s fix for stalled cycles is to turn permissions off.

This inverts agent-guardrails’ organizing principle. That page (and agents-never-do-alone) puts the bright lines at execution — a block list and human checkpoints, because you don’t trust the agent to infer limits. Auto-Company removes the enforcement layer and re-states the limits as prose the agent reads. It is also the exact trade run-claude-code-agents-24-hours warns against from the other direction: that source buys unattended autonomy with containment (sandbox + backups) precisely so permissions can be waived safely. Here the permissions are waived and the containment is the same host filesystem. Worth watching whether any incident reports land in the repo’s issues.

Personas as cognition

Layer 3 is the claim that role-prompting with named experts (“this is not ‘you are a generic developer’”) injects real mental models — Bezos working-backwards, Munger checklists and inversion, DHH’s majestic monolith — and thereby gives decisions “extreme business and engineering depth.” That is the persona half of reasoning-scaffolds, and it is asserted with no evaluation. The pointed counterweight arrived the same day: cc-thinking-skills packaged the same Munger/Goldratt/Boyd frameworks as skills, ran them through a placebo-controlled, replication-gated harness, and found no proven accuracy gain. Neither source settles the other — different task types, and Auto-Company’s output isn’t benchmark-shaped — but the spoke should stop treating persona injection as self-evidently load-bearing.

Where it sits

It is the breadth-and-time maximum of the unattended-agent thread: orchestrate-100-agents-claude-code scales one operator’s agents across tasks, run-claude-code-agents-24-hours scales one agent across a day, cognition-fable5-through-the-night shows a vendor agent running 8 hours on real work. Auto-Company removes the task and the horizon both, and asks what a loop with a charter converges on. The answer the repo actually supplies is in PROMPT.md: left alone, it converges on discussion, and the author had to hardcode phase transitions to stop it.

The engines are interchangeable by env var (ENGINE=claude|codex), with no fallback — the CLI is treated as a commodity executor whose file I/O, bash, and git access the orchestration layer inherits. Another instance of the agentic-coding-harness becoming substrate for something built on top of it.

loop-engineering · agent-memory · agent-guardrails · agent-orchestration · reasoning-scaffolds · run-claude-code-agents-24-hours · durable-agents · claude-md · zheyuan-kong · cc-thinking-skills