Product Agents Brief

Delegate a bounded task

Select a task with a clear input, output and reviewer. Research synthesis, draft acceptance criteria and analysis of non-sensitive feedback are useful starting points when their results can be checked.

Do not delegate responsibility for the product decision. State what the agent may inspect, what tools it may use and which actions require a person. Begin with read-only access where possible.

Task:
Allowed inputs:
Required output:
Tools and permissions:
Actions requiring approval:
Reviewer:
Time and cost budget:

Evaluate the work, not the confidence

Ask for evidence and uncertainty alongside conclusions. Check whether quoted examples exist, whether the analysis represents the full sample and whether suggested actions follow from the evidence.

Include incomplete information and conflicting requirements in the test set. The agent should ask for clarification or flag an unresolved tradeoff rather than manufacture certainty.

• Check factual statements against their sources.
• Separate observations from inferences.
• Inspect omitted or contradictory examples.
• Test tool failures and permission boundaries.
• Compare the result with a simpler manual workflow.

Keep a reviewable decision trail

Store the task brief, output, sources and reviewer decision together. When a proposal changes the roadmap, record the reason and the owner who accepted it.

Expand agent autonomy only when the team can explain its failures and recover from them. Keep a stop mechanism, usage limits and a clear owner for incidents.

Proposal:
Evidence:
Uncertainty:
Alternatives:
Human decision:
Decision owner:
Follow-up date: