Answer
How should subagents return results?
Condensed, with the evidence attached, and with what they could not establish stated as plainly as what they could.
Short, with references that let a claim be checked in seconds, and with the gaps stated. A summary that cannot be spot-checked has to be re-derived before anyone acts on it, which cancels the delegation.
The purpose of delegating was to keep material out of the main run, so the result has to be substantially smaller than what was read. A worker that returns everything it saw has moved the problem rather than solved it. Condensation is therefore the point rather than a courtesy, and the brief should say how short.
Condensation only works if it can be trusted, which is what the evidence requirement is for. Each substantive claim should carry the thing that supports it: the file and line, the quotation, the record identifier, the command and its output. This lets a reader check one claim in seconds and, more importantly, lets them check the one claim the next decision depends on rather than all of them.
Gaps have to be stated as prominently as findings. A worker that could not establish something will, by default, return what it did establish, and the omission leaves no trace. Requiring an explicit section — what I could not determine, and what I tried — converts an invisible deficiency into a visible one, and it is consistently the most useful part of a returned result.
The result should distinguish what was observed from what was inferred. A worker reading a configuration file and reporting the value has observed it; the same worker concluding what that value implies for behaviour has inferred it, and the two carry different reliability. Mixing them produces a report in which a solid finding and a reasonable guess are typographically identical, and the main run treats both as established.
Actions taken, if any were permitted, belong in the result separately from findings. A report that weaves what it discovered together with what it changed is difficult to review and easy to misread, and the changes are the part that matters most. Keeping them as a discrete list means the main run can verify them independently.
One structural point. The main run should treat the returned result as a claim rather than as established context, and marking it as such — this is what the search returned — changes how subsequent reasoning weights it. A result absorbed silently becomes indistinguishable from something the main run established itself, and its errors then propagate with the authority of first-hand knowledge.
The value of a delegated answer is set by how quickly it can be doubted, not by how confidently it is phrased.
Siddharth Sharma, Context Theory
Related questions
How short is short enough?
Short enough that reading it is obviously cheaper than doing the work, which is the whole justification. If a returned result takes as long to read as the task would have taken, the delegation has produced latency and nothing else. Specifying a length in the brief is cruder than it sounds and works better than asking for concision.
Should the subagent include its reasoning?
Its decisions, not its narrative. Which sources it chose and why, what it ruled out, where it had to interpret the question: these are checkable and useful. A retrospective account of how it thought is longer, unverifiable, and competes with the findings for the reader's attention.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | A returned result must be substantially smaller than the material read or the delegation has relocated the context problem rather than solved it, which makes a stated length requirement part of the brief rather than a stylistic preference. | Comparing the size of a returned result against the volume the delegated run consumed. |
| Response | Per-claim references allow a reader to check the single claim the next decision depends on rather than all of them, which is what makes condensation trustworthy without re-deriving the work. | Selecting the load-bearing claim in a returned result and attempting to verify it from the supplied reference alone. |
| Software | Observed values and inferred conclusions carry different reliability, and a report that renders them identically causes the main run to treat a reasonable guess as an established fact. | Marking each statement in a returned result as observed or inferred and checking whether the format distinguished them. |
| Constraint | A result absorbed into the main run without being marked as a returned claim becomes indistinguishable from first-hand knowledge, so its errors propagate with unearned authority. | Checking whether subsequent reasoning in the main run attributes the finding to the delegated run. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one