One-Person AI Company

Chat AI is good at tasks: answer this, rewrite that. An organization works differently. You hand someone an outcome and they decide what to research, what to do next, when to ask for help and when the work is done. The experiment tested whether that shape could be built from agents, with one human at the top.

The objective was deliberately real and measurable: a public pay-to-rank marketplace where companies bid for attention. The mission text stored in the run record asks the company to measure what every participant received relative to what they paid, identify who benefited most and least, explain why, and estimate the value captured by the platform itself.

Every run has its own monotonic event sequence. Fifty event types cover the whole life of the company: run lifecycle, consultations and tool calls, decisions made, applied or refused, task transitions, research activity, claims, verifications, snapshots, anomalies, process restarts, skipped heartbeats, missed checkpoints, human interventions and generated reports.

• 16 Prisma models: Run, Agent, Task, TaskAttempt, Event, Claim, Evidence, ClaimEvidence, Verification, HumanIntervention, ModelCall, Company, OutbidSnapshot, OutbidListing, RankChangeEpisode, BreakEvenScenario.
• Each event records who reported it and at what reporting level, so an agent’s account of its own work is never confused with a verified fact.
• Interim reports are preserved verbatim inside report_generated events; the run row holds the final one.
• Model calls store provider, model id, tokens and whether the cost was actually known. In this run it never was.
• Replay reads the same events through the same renderer as live, at up to 300x.

OpenClaw is reached through an OpenResponses-compatible gateway on the local machine. The backend sends the system instructions, the tool schemas and the previous response id; the gateway keeps the conversation and routes to its configured model (Claude Sonnet 5 with low thinking, per the run notes). The backend keeps nothing of the conversation itself.

• Used for: planning, tool-driven inspection of the dataset, structured decisions, interim and final synthesis.
• Not used: OpenClaw workspaces, files, browsers or its own routines. In this experiment it is a decision service, not a desktop agent.
• The single decision tool has a flat, object-rooted schema because the gateway’s tool bridge degraded a discriminated union into opaque strings during calibration. The backend still validates the strict union.
• Transport retries only on network errors and 502 / 503 / 504; never on a client timeout, because the gateway may be mid-turn spending real tokens.
• A mock gateway with MOCK labels exists for dry runs; nothing it produces can be mistaken for the real thing in the control room.

Read-only tools for the manager

get_platform_summary · list_companies · get_company · get_company_history · get_category_stats · list_rank_change_episodes · list_current_anomalies · list_existing_tasks · get_task_result · list_unresolved_questions · get_observation_metadata

The only write

submit_orchestration_decision

The manager can page through 1,140 companies, read a company’s bid, rank and click history, inspect rank-change episodes labelled CLEAN, CONFOUNDED or INSUFFICIENT_DATA, see its own tasks and their verdicts, and check how stale the observation is. It cannot write a row.

Rules injected into every research task

• Never click the marketplace’s outbound links: that would contaminate the click data being measured.
• No contacting founders, no purchases or sign-ups, no posting or commenting, no form submissions, no account changes.
• Public web only; public X posts only; never a human’s account, session or cookies.
• Attribution ladder with five levels, from a company-credited outcome down to UNKNOWN. UNKNOWN is a legitimate result.

There is no Telegram, git, file, email or browser-automation tool in this repository. Market observation is done by a human with a browser and a capture snippet, by policy: no scraping, no challenge bypass, human-scale volume.

How it stayed up for 24 hours

• Three macOS LaunchAgents with KeepAlive: the web app, the orchestrator worker and caffeinate to keep the machine awake.
• Five process_restarted events in the run: every restart is in the log, alongside everything else.
• The orchestrator’s launch command is itself the first emergency fix: node --env-file instead of a shell sourcing .env.
• Checkpoint alerts are local macOS notifications. There is no Telegram bot, no email, no daily briefing; a run summary is available on demand from the API.
• systemd units and a Caddyfile for a single VPS were written and never used; the gateway and the database are never proxied by design.

The working title of the article is “I Tried to Build a One-Person AI Company With Grok Bot + OpenClaw. Here’s What Actually Happened.” It is a local draft; there is no public link yet, so none is offered here.

• The business question: on a public pay-to-rank marketplace, is attention worth what companies pay for it? The measurable half (gross volume, distribution, cost per click) went to deterministic code. The unmeasurable half (what happened after the click) went to the researcher.
• Findings the company returned: about $182K in gross bids across 1,140 companies; the top 1 percent of spenders accounted for roughly 61 percent of spend; a 200x spread in cost per click across categories; visible audience mismatch; and commercial ROI marked UNKNOWN, because no company-credited outcome could be found publicly.
• The honest line the article turns on: I automated the brain, but not the eyes. The company reasoned overnight without me; it never once looked at the market on its own.

The company produced a usable interim report 33 minutes after launch and a final synthesis 70 seconds after the deadline. In between, it ran overnight without me. It also needed three emergency corrections, missed every observation checkpoint, spent tokens rewriting a report with no new evidence, and never had its run record formally closed.

Failures, in order

• The environment, not the model. Minutes after launch, the orchestrator was down because macOS refused a shell sourcing .env under launchd. The intelligence was fine; the plumbing was not.
• Synthetic data in a real run. Development seed data leaked into two read tools. OpenClaw did the rational thing with what it saw and opened a task about a company that does not exist. A human audit caught it; the cancellation is in the log.
• Autonomy stalled. Once research was declared sufficient, the design silently turned the remaining 23 hours into idle time. The fix had to be written during the run.
• Reports that got shorter. Each heartbeat consultation rewrote the report with no new evidence: 9,491 to 7,338 to 6,714 characters. The system was productive-looking and useless, at a cost.
• The eyes were never automated. All four observation checkpoints were missed because fresh market capture stays human-attended by policy. The company had a brain and hands, but I was still the eyes.
• Cut-off thinking. 23 of 41 consultations hit the 6,144-token output cap and ended as truncated. They were logged as such and never treated as decisions.

The hard part

Building agents that act was the easy half. The hard part was building the thing that decides whether to believe them: a verifier that grades every claim, an event log that cannot be edited by hand, a UI that refuses to display anything it cannot trace, and a rule that UNKNOWN is a legitimate answer. Most of the mid-run fixes were about that layer, not about the model.

Agent contract

1. Objective
   One mission with a hierarchy: business result first, experiment second, proof of work third.
2. Orchestrator loop
   A TypeScript worker (no framework, no queue): poll, guard, trigger, consult, validate, journal.
3. OpenClaw gateway
   OpenResponses-compatible /v1/responses endpoint on loopback, bearer token, transport-only retries, response-id continuity.
4. Task inbox
   Five narrow HTTP endpoints behind a worker token: peek, claim, events, result, fail. Claims are atomic and exclusive.
5. Verification
   Deterministic checks per claim; an optional Claude semantic judge behind an environment flag, not enabled in the run.
6. PostgreSQL
   16 models, 3 migrations: runs, tasks, attempts, events, claims, evidence, verifications, interventions, model calls, market data.
7. Control room
   Four views over one API: live (SSE), results, evidence, replay. Server-derived topology state.

I defined the mission hierarchy and the epistemic classes, designed the decision contract and the invariants, wrote the orchestrator loop, the OpenClaw adapter, the task inbox, the verifier, the event schema and the four-view control room, ran the calibration and dry runs, launched and supervised the official run, wrote the three corrections while it was live, and wrote the article afterwards. The research itself was done by the agents; the market observations were captured by me.

Next

• Milestone 2: verify what happens after the click with founder-provided, attributable evidence, so ROI can move from UNKNOWN to a graded claim.
• Automate the eyes: scheduled, policy-compliant observation so checkpoints stop depending on me.
• Enable the semantic judge and report deterministic and semantic verdicts side by side.
• Close the run properly: the record still says ACTIVE.