Write a Multi-Agent Delegation Brief That Prevents Rework

Start with the deliverable

Describe a result someone can review. For a hypothetical signup bug, ask the worker to identify why an existing email address produces a server error and propose a focused fix. Include the expected response, a reproduction path, and the files already known to be relevant.

State whether the worker should change code. An investigation and an implementation have different finish conditions. Mixing them can leave you with an unrequested rewrite when you needed a diagnosis, or a long explanation when you needed a working patch.

Include the context that changes decisions

Pass the constraints that matter to this assignment: supported behavior, nearby work in progress, permitted tools, and anything the customer has already decided. Link to stable artifacts such as a ticket or specification instead of relying on phrases like "the thing we discussed earlier."

A useful context packet includes the current revision. If the worker reads yesterday's contract or an older API shape, its answer may be internally consistent and still wrong for the current task. Ask it to identify any mismatch before making assumptions.

Define ownership and stopping points

For code work, list the files or area the worker owns and name the shared files that require coordination. For document work, identify the section it may edit. Give the worker a way to report that the assignment crosses that boundary.

Set a sensible stopping rule. A worker that cannot reproduce the signup error should return the attempted steps and missing information. It should not keep inventing different bugs to fix until something looks productive. A clear blocked result can save more time than an unrelated patch.

Ask for evidence in the return message

Request the outcome, relevant artifact paths, checks performed, and remaining questions. Keep the evidence proportional to the task. One small fix may need a reproduction result and one targeted check, while a broader migration needs a more substantial record.

Here is a brief structure you can adapt without turning every assignment into a long document:

Read the finished brief from the worker's point of view. If a sentence depends on context it won't receive, add that context or remove the dependency. The goal is to make the next action clear enough that the worker can spend its effort on the task.

• Outcome: what should exist when the worker finishes.
• Inputs: files, records, revisions, and examples to use.
• Scope: what the worker owns and what needs coordination.
• Evidence: how another person can check the result.
• Escalation: what uncertainty or failure should stop the work.

Sources & further reading

• Passing context between agents