Cognition co-founder and chief product officer Walden Yan explains the company’s internal agent workflow in this 14-minute AI Engineer talk.

Small teams own the whole problem · 0:34

  • Yan reports 40 times as many merged pull requests over a year while engineering headcount grew 2.7 times.
  • He says code quality held or improved; merged-request counts alone do not measure delivered value, reliability or engineer productivity.
  • Two- or three-engineer pods own customer feedback, product choices, quality assurance and their work queue, with agents handling implementation.
  • Some changes run from bug report through merge and monitoring automatically; human judgment enters through escalations.

Return evidence where the work starts · 3:31

  • A customer sends a bug video in Slack; the team forwards it to Devin and receives a pull request plus a recording of the fix.
  • Engineers can inspect the result and relay evidence to the customer without switching into an editor for every change.
  • Yan’s example depends on the test environment underneath the chat interface, not just connecting an agent to Slack.

Make the whole application runnable in isolation · 4:28

  • Cognition migrated services to Docker Compose and supplied local mocks for third-party services, according to Yan.
  • Each agent gets a virtual machine with the application and test dependencies, avoiding live-service access for these tests.
  • VM images rebuild from current code every hour; Devin is the first responder when those rebuilds fail.
  • Agents maintain testing skills as the code changes, including multi-account voice and video-call checks.
  • Centralized execution can reduce secrets on laptops, but the talk’s security claim depends on how those cloud machines and credentials are controlled.

Let agents improve the workflow itself · 7:05

  • Agent traces can reveal wasted time, excess token use, repeated work that should become scripts and missing tests.
  • Yan describes a daily Slack channel for proposed improvements, such as cutting build time or catching bugs earlier with lint rules.
  • An ongoing performance job could add telemetry, establish a baseline, create alerts and report progress on P99 latency, the slowest one percent’s response-time boundary.
  • He proposes turning measurable goals into persistent work; choosing the right metric and checking regressions remain engineering responsibilities.

Persistent jobs still need resources and review · 9:49

  • Cognition allocated GPUs to an agent for roughly a week of inference tuning; promising strategies went to an engineer for final approval.
  • Another agent monitors iOS code and customer feedback to maintain an Android counterpart, with automated visual and functional checks.
  • Yan describes 44 days of Android development, some days with no human messages and upwards of 40–50 merged pull requests per day.
  • At recording time, that Android app was described as working and approaching release, not independently validated or already generally available.

“It’s not just you purchase a coding agent, you got to go change the way you build.” — Walden Yan, 4:37–4:40