Devin Bots turn engineers into managers of AI coworkers
Key points
Key takeaways from an MTS interview with Cognition co-founder Walden Yan on the launch of Devin Bots (November 2026):
Jobs, not tasks. Devin Bots are designed as durable, proactive co-workers that own a job — running a release process or acting as chief of staff — rather than completing individual tasks; one beta tester runs her gaming company with 20 specialised bots handling everything from UI design to release management, and her weekly merged PRs reportedly grew from around 50 to 1,900.
Devin builds Devin. Yan says he struggles to find code at Cognition not written fully autonomously by Devin; the whole development process has moved onto the product, and employees increasingly work from iPads running Devin in the cloud.
Autonomous Android parity. Cognition's Android app was built by giving a single bot one instruction — maintain full parity with the iOS app — and checking in occasionally while the bot merged thousands of PRs, running tests and pixel-level comparisons itself.
Decisions as the human interface. Because agents run so many threads in parallel, Devin surfaces only the decision points that genuinely need a person and auto-merges low-risk changes, cutting context-switching overhead.
4x leverage. Internally, the ratio of merged PRs to messages sent to bots rose about 4x since bots were adopted, and bots now commonly manage other bots.
Proactivity needs context. The groundwork came from the acquisition of Poke, the iMessage agent that could text reminders; its context-building — linking GitHub, Slack and a user's history to anticipate needs — is what makes Devin Bots proactive, including flagging related work by other people or submitting cross-cutting fixes.
Beyond vibe coding. Yan argues the industry is past vibe coding: people will build and maintain professional, custom tools with agents that fill in the gaps, and his main tip for users is to be far more ambitious — "throw it all at it" — because bots work on many threads in parallel and compound learning over time.
Read more: YouTube — MTS
