- Walden Yan of Cognition argued that engineers should give agents measurable goals to work toward over long periods, instead of single tasks.
- Humans stay in the loop at the end: Yan described engineers approving the best strategies an agent found and ruling on only the most important decisions.
Walden Yan, of the AI coding company Cognition, argued on the AI Engineer show that engineering is moving from handing agents tasks to handing them jobs. The episode was published on 18 November 2026. Yan described engineers setting a measurable goal and approving key results. Calling that skill judging is our reading, not his wording.
What was said
Yan described "moving from the world where you're giving tasks to your agents to giving jobs to your agents". His example was P99 latency. That is the response time that the slowest one percent of requests exceed. The agent would add measurement, set up alerts and send daily updates on progress.
He said: "We don't even want you to be prompting your agents most of the time". Engineers give goals instead. He also urged them to find a way to measure anything they think is not measurable, such as click latency or onboarding retention.
Cognition pointed an agent at throughput on new NVIDIA GPUs for about a week. A human engineer gave final approval to the best strategies it found.
In a second case, an engineer told Devin, Cognition's agent, to keep an Android app level with the iOS app. The iOS app was the reference. Permanent hooks watched its code, and the agent read Slack feedback channels. Live QA compared the two apps pixel for pixel.
Over 44 days, Yan said, this produced a working application. On most days no human sent the agent a message. On those days it merged upward of 40 to 50 pull requests a day.
Why it matters
Our reading: if the agent does the making, the scarce skill is the brief and the check.
Yan described the Android job as producing a working app. If that holds, it would be because the iOS app and pixel-level QA gave the agent a checkable standard.
Choosing a metric that reflects what the business wants is the hard part. Teams that cannot yet measure their goals have homework to do before agents can help.
The other side
Yan spoke as a vendor describing his own company's tools. The Android app was not yet released at the time of the talk. The excerpts show no outside check of its quality. Yan said only that it was QA-tested against the iOS app.
The talk also leaves two questions open. One is how people learn to judge agent work without doing the making. The other is how a team spots a metric that is being gamed.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: 40x More Merged PRs With 2.7x the Engineers — Walden Yan, Cognition (2026-11-18)





