Most coding agents are evaluated at the moment they answer. Software teams discover whether an answer was useful much later: a test fails, a reviewer requests changes, the pull request stalls, or the code ships. This talk shows how to turn those delayed signals into a continuous evaluation system. We instrument AI conversations, tool calls, diffs, commits, CI runs, review comments, and merge outcomes, then join them into one trace from prompt to production.
Using a working implementation, I’ll walk through the event model, correlation logic, and queries that separate accepted work from rework and repeated failure. I’ll show the hard parts: carrying trace context across unrelated tools, distinguishing flaky CI from agent errors, protecting developer privacy, and avoiding vanity metrics. Finally, we’ll turn recurring corrections into human-approved candidate skills and test whether they improve first-pass CI, review rounds, and time-to-merge.
You will leave with a product-agnostic blueprint for outcome-based evaluation of coding agents.
This talk has been presented at AI Coding Summit Berlin, check out the latest edition of this Tech Conference.
