AIエージェント pipeline: the real bottleneck is queue time

The account begins with a simple goal: figure out whether an AI-agent pipeline that opens pull requests from GitHub Issues actually got faster. The first attempt, summing time from Issue creation to close, failed because that span mixes detection, assignment, implementation, CI, review, merge, and close. The author then tried to split waiting from implementation using the first commit's timestamp, but the measured implementation time came out to a median of one minute, because this pipeline commits once at the end and immediately opens the pull request.

Treating the post-PR period as pure waiting was also wrong. In many pull requests the last commit came after the PR was opened, by gaps ranging from tens of minutes to three and a half hours, because review feedback generates new commits. The author concludes that only external timing records, capturing dispatch, implementation completion, and handoff, can separate the intervals, and that the first fix should be adding measurement, not improving the scheduler.

The author's four-way split of waiting time and the proposed response, a dispatcher that only assigns and a worker released at handoff, point to the same reading: the constraint appears to sit in flow management rather than model performance, though this reflects a single operating environment and the method's usefulness hinges on whether the added timing points are actually instrumented.