Browserless fetch snapshot for agents, crawlers, and cheap verification tools. The interactive React app still hydrates this route for humans.

Durable Agent Work Is More Than A Worker - AIIDIOTSAI Field Notes

The OfficeMac/Superbox export proved that durable workers are necessary but incomplete. The useful loop needs purpose, sensed state, decision records, verification, reporting, and a lesson that the next agent can execute.

Lesson: A durable worker is only the chassis. The reusable pattern is Purpose -> Sense -> Decide -> Execute -> Verify -> Record -> Report -> Learn.

Source and boundary This note comes from a sanitized MacBook field-report export produced after the Superbox/media-control work. The private export stays private. This public version excludes private messages, handles, account details, credentials, stream URLs, protected media internals, device secrets, and external-send claims. The safe lesson is still valuable: the build confirmed the durable-worker foundation, then exposed the next layer AIIdiots has to teach. What the MacBook taught Chat is a good front door. It is not a runtime. The Superbox work needed durable state outside the transcript: a local work queue, receipts, intent notes, git state, and a way for the human to know whether the agent was active, blocked, stale, or done. That part matched AIIdiots doctrine. But durable mechanics alone were not enough. A long-running agent can still become useless if it reports vague progress, hides uncertainty, treats local commits as publication, or dumps private receipts into a public-facing surface. The missing layer Intent beats telemetry. A useful report says what the agent intends to do, why, what state it sees, how it will prove progress, and when the human is actually needed. State must be obvious. Every loop should be active, idle/no-change, blocked, waiting on human, stale/needs attention, or done/closed. Local git is not publication. Agents must distinguish local commit, remote push, GitHub visibility, tracker mirror, and production deploy. Durable runs need isolated context. The chat controls the run, but the run needs its own folder, state files, receipts, and closeout digest. Riffs need a queue. Raw ideas, execution receipts, planning backlog, and publication state are different objects. Recommendations need confidence language. Content confidence, source freshness, device route confidence, selector/OCR confidence, and playback verification are separate claims. Observability needs privacy filters. A dashboard that blindly shows the newest receipt can leak private context. Catalog first, device second. For media-style work, decide what is worth watching from source/catalog/taste data before crawling devices to verify or play it. The reusable loop That loop is the difference between “I have a worker running” and “I can trust this system to keep learning while I sleep.” What changed in AIIdiots This report becomes the first public paired field-report output under the current rule: Human version: this field note. Agent version: /lessons/superbox-agent-loop.json. Future agents should use the paired lesson before claiming durable work is under control. If they cannot name the purpose, state, queue, gates, confidence language, receipt, and lesson output, they do not yet have the loop.

Canonical route: https://aiidiots.ai/notes/durable-agent-work-is-more-than-a-worker