Q21 of 60 — CCA Practice Exam

Multi-Agent Research System

Your multi-agent research system uses context compaction between turns: after each subagent completes a research subtask, the orchestrator summarizes that turn's history before issuing the next request to Claude. After deploying a 128,000-token task budget, your logs show that some research jobs consistently overshoot the intended budget — tasks that should finish within 80,000 tokens are consuming upward of 115,000 before gracefully stopping. Examining the orchestration code, you find that task_budget is set correctly on the first request of each job, but the remaining field is omitted on every subsequent request after compaction. What is the most likely explanation for the overrun?

A. The server resets the budget countdown to total on each post-compaction request because it has no memory of prior spend, so Claude self-regulates against the full 128,000 tokens at the start of every compacted turn instead of the true remainder.

B. Compaction itself consumes tokens counted against max_tokens but not against task_budget, so summarized turns are excluded from the countdown regardless of whether remaining is supplied, creating an untracked token sink.

C. The effort: "high" setting overrides the task budget during compacted turns, causing Claude to ignore the countdown until the final turn of the loop, at which point graceful stopping finally engages.

D. Omitting remaining causes the server to treat the budget type as unbounded, disabling the graceful stopping behavior entirely for all turns that follow the first compaction boundary in the job.