Build an autonomous daily desk, not an autonomous company

The result

Each working day, a small operations desk reads an approved queue, prepares a limited number of internal drafts and puts them in a human review inbox. No work means a recorded no_work result, not manufactured activity.

A useful first example is an inquiry-preparation desk: read a few incoming project requests, gather the approved facts needed for a reply, and prepare drafts. It does not send emails, set prices, commit delivery dates, change permissions or spend money.

This is one workflow, not a blueprint for running an entire organization.

The smallest useful contract

Approved queue → bounded preparation → internal drafts → human review
                       ↓
                completion record

Download queue.mjs. It demonstrates a capped, deduplicated batch that can only return draft proposals. It has no model, scheduler, inbox connection or send capability. The sample's three-item default is an illustrative work budget, not a measured optimum.

Build the desk

1. Select a queue and a result. Pick one input source and one deliverable. Write down what a useful draft contains and which missing facts require the desk to stop. Use synthetic requests until permissions and data handling have been checked.

2. Grant the minimum capabilities. Reading the approved queue and writing internal drafts should be enough. Keep outward actions outside the worker's credentials. A sentence in a prompt is not a substitute for removing a send or payment capability.

3. Bound every run. Cap items, execution time and model spend. Use an atomic run claim or lease to prevent overlap, plus durable item IDs to avoid processing the same work again on the next day. The in-memory sample only deduplicates within one batch.

4. Keep the source attached. A reviewer should see the original request, relevant evidence, unresolved questions and proposed response. Untrusted request text is source material, not permission to change the desk's instructions.

5. Record actual completion. Store the run ID, input cursor, output references, actual usage and final state. A configured schedule is not a completed run. Mark no_work, completed, failed and needs_access separately.

6. Add a simple external check. Compare the last completed run with the expected schedule and report a missed deadline. Monitor the delivered drafts as well as job execution. A process can run successfully and still produce nothing useful.

7. Improve from human decisions. Track accepted drafts, correction effort and rejected suggestions. Change the workflow only when that evidence identifies a specific problem. Do not respond to poor output by adding more agents automatically.

What running one every day taught us

These points come from a daily reconciliation run that has kept several projects' notes and task boards current.

Check cheaply before reading anything. Ask what changed since the last run. Read only those changes. A run that rereads everything every day costs the most on the days it has the least to do.

Stay quiet when there is nothing to act on. A daily message that says "nothing to report" teaches people to ignore the desk. Silence on a quiet day is the correct output, and the completion record still shows the run happened.

Write the boundary as two lists. What the run may do alone: reconcile notes, check that work marked done has evidence, merge a ready internal change after its required checks. What it never does: merge product changes or deploy. The second list matters more.

The desk is the backstop. Whoever does a piece of work records it when they finish. The daily run catches what was missed and reconciles conflicts. If the desk is the only thing keeping records current, the records are a day old.

Give the run a name and a time. A named heartbeat at a fixed hour makes a missed run visible as an absence. Pair it with the external check in step 6.

Check claims against evidence. "Done" with no commit, file or link behind it goes back as open. This is the most useful thing the desk does.

Failure rehearsal

Before turning on a schedule, simulate a revoked credential, an empty queue, a duplicate event, an overlong request, a model timeout and a reviewer who is unavailable. The desk must leave recoverable state and must never turn missing approval into permission.

A retry is safe only when it cannot duplicate a consequential action. Keep writes to the draft store idempotent, and never put an email send inside an automatic retry loop.

Evidence and maintenance

Offline tests cover capped batch size, within-batch deduplication, empty input and draft-only output. They do not prove scheduled reliability or an autonomous organization. A real pilot must record completed runs, accepted outputs, human correction time and actual cost before a stronger claim is appropriate.

Review permissions and missed-run checks regularly. Stop the desk when its drafts consistently cost more effort to correct than they save.

Agent implementation instructions.