Help › Performance · version 8 ·

AI agent performance: how long runs take

Conductor projects. Performance (Log:Performance) shows how long your agents' runs take. Every run a coder (or another agent) finishes leaves its timings, step by step: launch, reading, setup, build, tests, merge, docs, checkpoint and deploy. It also records where the run ran: in the cloud, on a laptop, or dispatched.

Read the summary

The page is a query. It opens on the last 7 days with a summary line: the number of runs, the median minutes per step, the fixed time (what every run pays: launch, setup) against the time per task, and each runner apart. Below it is one row per run.

Find what slows your agents down

  • "Are tests the slow part?" Compare the median tests minutes with build. If tests take most of a run, it is worth splitting the suite or running fewer of them per task.
  • "Is the laptop slower than the cloud?" Pick one runner, then the other, over the same week, and compare the medians.
  • "Why did yesterday's runs take twice as long?" Set the range to yesterday and read the rows step by step. A long setup or launch usually points at the environment, not the task.
  • "Do small tasks cost too much?" When fixed time is close to the time per task, batching small tasks into one run saves the most.

How long it is kept

Entries stay on the project for 7 days. Every night, whole days (UTC) older than that move to one file in your project's Files, folder _archive/, named history-<first day>--<last day>.jsonl.gz. A query that reaches further back names the files that cover it. How the archive works, and when old files are deleted: Project history.

Updated by razie · version 8