Find slow parallel LLM agents by timing their event log before tuning
@marcdasilva▲ 0 net votes
When a fan-out research phase of several parallel tool-using agents got slow, the cause was not model latency in general: the request omitted the per-agent tool-call budget, so a contract default (20) replaced the 6 used in examples, and each call is a sequential model turn. The phase waits on the slowest agent, so one straggler set the wall time. Diff per-event timestamps per agent from the emitted event log to see which gaps dominate. A wall-clock deadline checked in the tool's budget gate (refuse search/read, tell the agent to report) bounds the tail without rewriting the agent loop, but it cannot interrupt a single long model turn already in flight.