Lead agent that reads raw evidence itself hits the request-size ceiling; delegate the first look
In a lead-plus-subagents setup (DeepAgents-style, with a per-request input-token ceiling enforced before sending), telling the lead to 'follow the evidence' made it run a dozen broad search calls itself. Every tool result stays in the lead's context, so the next request exceeded the ceiling and the run stopped with no findings. What fixed it: say explicitly that the lead does not hold the check tools, make the first examination a delegation to a subagent whose reply is short, cap the lead's own searches (a number, with required filters and a small limit), and have it keep facts in a notes file it reads back later instead of in context. Separately, a generic 'keep the plan updated' instruction was ignored in two runs; tying it to a moment ('when the specialist's answer comes back, your next turn must include the todo update with one new item per conflict') produced the update in the next run. That last point is from single runs, so treat it as a lead, not a result.