Devin’s 80% Moment: Background Agents, 7x PRs, & End of Hand-Held Coding — Walden Yan & Cole Murray

TL;DR

• 1

  Devin’s usage hit an 80% tipping point — Walden says Devin-authored code in Cognition’s repos rose from 16% in January to 80% in March, while merged PR volume grew roughly 7x in 2-3 months.

• 2

  The real challenge is testing, not computer use — Clicking UI elements is the easy part; the hard part is reasoning through how to run multi-service apps, enable flags, get the right permissions, and verify cross-stack changes end to end.

• 3

  Out-of-the-box agents are safer, but much harder to build — Running the “brain” outside the sandbox protects secrets and supports cleaner permission boundaries, but forces you to manage state, orchestration, and more complex infra.

• 4

  Most companies still aren’t actually ready for autonomous coding — Repo setup remains a bottleneck because many teams still rely on tribal knowledge like “go ask Bob for the secrets,” which breaks when an agent needs to boot and test a repo on its own.

• 5

  AI coding quality now depends on guardrails as much as model intelligence — The guests call out repeated failure modes like backward-compatibility hacks, untyped tuples, getattr reward-hacking in Python, and codebase “slop” spreading from weak patterns unless linting and cleanup are enforced.

• 6

  The fastest-growing use cases are outside classic engineering — SRE triage, customer support, PM-driven bug fixes in Slack, and internal knowledge workflows are all becoming agent-native because the value comes from integrating code, logs, docs, tickets, and chat in one loop.