How to Delegate Work to AI Agents

How do you write an outcome-focused agent goal?

State what should exist or be true when the job is finished. “Research competitors” is an activity; “produce a cited comparison of five competitors with pricing, audience, and unresolved questions” is an outcome that can be reviewed.

Include the intended reader or downstream use. The same research should look different when it supports a product decision, a sales call, or a public article.

What context should you give an AI agent?

Give the smallest set of authoritative context that makes the job unambiguous: relevant documents, current data, definitions, examples, and the system of record. More context is not automatically better when it contains stale or conflicting material.

• The current source of truth and its date or version.
• Definitions for business-specific terms.
• Examples of acceptable and unacceptable outputs.
• Known constraints, exclusions, and edge cases.
• Where the finished artifact should be stored.

How do you define agent acceptance criteria?

Acceptance criteria should be observable. Specify required fields, tests, citations, file formats, limits, and review checks. Avoid criteria such as “make it good” unless you also define what good means for this job.

For code, require relevant tests and a concise change summary. For research, require source links, source dates, and explicit uncertainty. For content, require factual support and a named audience rather than a word-count target alone.

How should you limit agent tools and permissions?

Grant only the tools and credentials needed for the goal. Separate read access from write access, use scoped credentials, and avoid giving a research task permission to publish or delete.

Tool descriptions should explain both capability and boundary. Similar or overlapping tools increase selection mistakes, so start with a small, well-defined set.

Which agent actions need approval?

Require approval before actions that create external commitments or material risk. Publishing, sending, purchasing, deleting, changing production, and exposing private data are common checkpoints.

Do not approve from a vague summary. Review the proposed action, affected target, supporting evidence, and rollback plan when one exists.

How do you review an agent's result?

Review the artifact against the original acceptance criteria, then sample the underlying evidence. Check what the agent did not complete, where it made assumptions, and whether any source or dependency changed during the run.

A failed review should produce a narrower correction with new evidence or criteria. Repeating the same broad instruction usually repeats the same ambiguity.