Delegating work to AI agents: write the handoff first

Apply the fit / stretch / trap screen to the audience job before assigning it. Ask whether the task follows a familiar pattern in public writing and whether wrong consent will be caught. Check whether you are using the agent to prepare a decision or letting it make that decision for you.

• Fit: a familiar task with enough context and cheap or checked errors. Delegate a summary of an already reconciled audience record.
• Stretch: a task that needs deliberate verification. Give an audience reconciliation a separate check against both consent sources.
• Trap: missing grounding or consequential errors that pass unchecked. Keep a consent-rule change with the person who owns that decision.

Write “what stays with the owner,” the decisions you keep, beside your label. Follow the director role in the audience repair: let the agent prepare the list while you retain the consent rule and send approval. Keep a trap with the person; tighten the check before dispatching a stretch.

Open the charter, the standing job description that sets an agent's scope and authority. Write this job's outcome inside that scope using the lab's decompose-before-dispatch rule: split the assignment into a bounded checklist whose items can each pass or fail. Use this guide's six lines to make the handoff checkable before it leaves you.

Replace “work on the campaign” with “return an audience checked against both consent sources, with conflicts excluded and sending held for approval.” Use the false completion reports as your rejection case: require an inspectable audience before crediting work on that outcome.

Name both the ecommerce platform and the legacy email platform in your record line. Use the audience repair to inspect what the tool actually reads: its final consent check discarded the legacy source even though merged consent was available.

Carry both consent-source locations and the rule into the handoff, then require the worker to identify the records it used. Check the last filter as well as the inputs; the audience tool's extra store-only check recreated the defect on every run.

Write what the agent never does in the limits line: for the audience job, never send before owner approval or include an explicit opt-out. Attach each limit to a gate, a checkpoint with authority to stop the action. Name the owner who can release the send and the consent evidence they need.

Use Bizzy's pre-commit IP gate on the Example team as the implementation model: the client-code boundary blocks internal org files at commit time. Try the forbidden action in a test copy before calling your own limit enforced. Mark the remaining written limits honestly; the constitution account distinguishes written rules from those already wired.

Fill the check line with a test that would fail the audience tool's incomplete reconciliation. Give its owner the full source list before the worker runs. Keep “check” and “who holds it” on separate lines so the worker cannot silently accept its own consent assumptions.

Separate technical readiness from commercial acceptance using the built-right and will-it-earn gates. Assign the commercial veto to a different owner before commissioning a product. For a high-stakes release, use an out-of-family reviewer, an agent from a different model family, as in the held-release account. Require a reproducible finding before clearing a dispute with that reviewer.

Put the receipt's destination in the last line before dispatching the job. For a site release, require the deployed URL and its observed browser result there, using the blank-site account to decide what belongs in the evidence.

Open the browser evidence when it arrives and compare it with your page's outcome line. Reject a tests-only receipt for a page-rendering claim: the lab's security setting blocked the page while its tests stayed green. Save the blank-page observation beside the corrected one so the next owner can see why you held the job.

The audience repair before send approval

Reconstruct the handoff from the audience account and the map of the company brain, the shared context and permissions agents read. Use the map for the seat sequence and the account for what failed and was repaired.

1. Owner's ask. I needed the campaign's email audience built under the current-consent rule. Keep that request separate from permission to send, as the map's approval hold does.
2. Owner agent's decomposition. Give the owner agent, the agent responsible for a business within its charter, the map's split: CMO prepares the audience; Counsel checks consent; COO reconciles and verifies. Carry both source systems into that assignment. The failed tool checked only the ecommerce platform and reported 4 sendable contacts.
3. Gate that held the send. In the map, the send waits for the owner's approval. Keep that hold in your handoff through the repair. A separate agent questioned the incomplete source in the account; repeating the worker's spot checks would have missed it.
4. Fix. The decision I kept was which consent rule the repair had to implement. The tool was changed to use consent from either platform while excluding explicit opt-outs. Cross-source conflicts were repaired and the update path changed so the sources stopped overwriting each other.
5. Receipt. Open the published reconciliation artifact to check the merged-consent repair. Read about 250 reachable as the reconciled pool; apply the remaining send filters before treating it as a final audience. Keep the owner's approval beside that result before any send.

• An activity in the outcome line. Replace “worked on the page” with a result someone can inspect; the blank page reported complete shows what the activity report can hide.
• A source name without a source check. Inspect the audience tool's final filter; the store-only repeat check threw away consent that the merged record already held.
• A limit left in the prompt. Locate the stopping point before dispatch; the constitution still had written rules awaiting enforcement.
• A reviewer whose findings end the argument. Reproduce a disputed failure before rewriting the job; the outside audit included false build-failure findings.
• A receipt location nobody opens. Read the browser evidence before closing a site job; the green suite over a blank site could not answer whether the page appeared.

Pick one job you would hand off tomorrow and mark it fit, stretch or trap. Use the audience repair as the pattern, then replace these brackets with your own job's facts. Keep a trap with you; give a stretch the extra check it needs before dispatch.

Read your six lines as the checker from the audience account: would they expose a missing legacy source before a send? Mark any unanswered line and resolve it before tomorrow's dispatch. Save the completed handoff beside the job so the worker and checker receive the same consent rule.

Run the handed-off job with the daily operations checklist. Test its checks with the lab's failure drills.