The Five-Field Handoff
Delegating to a sub-agent fails on missing context, not on missing capability. Five fields fix most of it.
Sub-agent delegation usually fails because the brief omits why the task exists, so the agent optimises the wrong thing. Five fields — task and why, context, output, limits, verification — cover most failures, and the why field is the one that changes results.
Prerequisites
An orchestrating agent that can spawn sub-agents, and at least one task big enough to be worth delegating.
Steps
1. Write the task with its reason attached. One field, two parts. The reason is not decoration — a sub-agent that knows why can make the dozens of small judgement calls the brief did not anticipate.
Task + why: Research what content in this niche gets traction. Without it we will write what we enjoy rather than what anyone searches for, which is how the previous attempt died.
2. Give context the sub-agent cannot derive. What exists already, what was tried, what the constraints are. Sub-agents start with no history. Most bad output is a reasonable answer to a question that was missing three facts.
3. Specify the output shape. Format, length, structure, and what not to produce. Vague output requests return essays; specific ones return usable material.
Output: markdown, max 1800 words, no preamble. Top 15 items, each with evidence of demand and a competition estimate.
4. Set limits explicitly. Confidence marking, source requirements, what to do when uncertain, and the boundary of the task. The most valuable limit is permission to fail: if you cannot find it, write that you could not, do not fill the gap.
Without that permission, an agent optimising for a complete-looking answer will produce a complete-looking answer.
5. Ask for verification in the reply. Make the sub-agent state its own weak points: how many independent sources, which claims rest on one, and the single thing it is least sure about.
This is the highest-yield field per word. It converts an over-confident report into a usable one, and it costs one sentence in the brief.
6. Send the brief to more than one agent when the answer matters. Two agents on the same brief, working independently, produce a cheap check: where they agree the finding is probably solid, where they diverge is where a single agent would have handed you false confidence.
This costs roughly double and is worth it exactly when the output feeds an irreversible decision — which is a small fraction of tasks. Using it everywhere is how delegation stops paying for itself.
Verify
Read the returned work and ask what you would have had to check by hand. If the answer includes "whether these numbers are real", the limits field was too weak.
Then check for the failure mode that a fan-out hides: partial success reported as success. Any parallel dispatch must report X of Y completed, with failed branches named. A summary that silently covers 4 of 5 branches is worse than one that covers 4 and says so.
One more check worth running once: hand the same brief to a colleague and ask what they would need to ask you before starting. Every question they raise is a field the brief is missing, and it is far cheaper to find them that way than in a returned deliverable.
Troubleshooting
Output is generic. The why field is missing or thin. An agent that does not know the purpose defaults to the average answer for the topic.
The sub-agent invents numbers. Add explicit marking — measured, derived, estimated — and an instruction that unfindable is a valid result. Fabrication is usually a response to a brief that made "I don't know" feel unacceptable.
Two sub-agents return contradictory findings. Good. That is the signal a single agent would have hidden; reconcile it in the open rather than picking the more confident one.