Research is the pipeline's first optional detour, and most efforts never take it. It exists for one situation: the grill reaches a decision that turns on a fact nobody in the room can state with confidence — how a third-party API actually behaves, what a library really supports, where a platform's limits sit, what something costs.
The tell is almost always the same sentence: "I think X probably supports this." That "probably" is the signal. A decision resting on a probably is a decision resting on a guess, and guesses made during a grill harden into requirements downstream. So the grill pauses and the fact gets verified before it can do any damage.
How it runs
Research runs agent-solo — this is the one part of the early pipeline a human doesn't need to sit through. The agent goes to primary sources: the vendor's own documentation, the specification, the actual API surface, not a summary of a summary. The findings are written up as a short, cited document and committed to the repo, so the answer is durable and auditable rather than trapped in a chat log that evaporates when the session ends.
The step isn't finished when the document exists. It's finished when the answer is read back into the grill. Research is always in service of a decision that's currently stalled; the whole point is to unstick it, so the finding has to return to the conversation that needed it.
What gets settled here
Exactly one thing: the fact. Research doesn't make the decision — it removes the uncertainty that was blocking the decision. The grill still owns the choice; research just makes sure the choice is made against something real.
For the genuinely hard cases there's a heavier engine available. When a question is contested or spans many sources, the step can reach for a web-scale fan-out-and-verify research harness underneath — but the output contract never changes: primary sources, cited, written down.
Built on
This step is built on Matt Pocock's research skill. Its output contract — cited findings drawn from primary sources — holds even when a heavier research engine does the digging underneath for a contested question.