Devin 1.2 adds smarter repo reasoning and voice clips

Devin just picked up your Slack voice memos.

Cognition, the Scott Wu-led startup behind what it calls the world’s first AI engineer, has rolled out Devin 1.2. The release shifts focus from raw benchmark bravado to the messy, practical reality of working inside enterprise codebases.

Voice memos and repository awareness

The marquee upgrade in Devin 1.2 is sharper in-context reasoning across repositories. Instead of treating tasks in isolation, Devin is now better at grasping an entire repository's architecture, figuring out which files actually matter, and reusing existing patterns instead of reinventing the wheel with bloated abstractions.

Then there is the Slack integration. Users can now simply tag Devin in Slack, hit "Record audio clip," and speak their instructions or feedback out loud. Devin breaks down the voice prompt into steps and gets to work across its sandboxed shell, code editor, and browser. It eliminates the friction of typing out structured multi-paragraph prompts on a keyboard.

Enterprise tooling and the reliability hurdle

Cognition also addressed day-to-day workflow friction with machine snapshots. If you log into a service through Devin’s browser during setup, the system preserves cookies for future sessions, solving headaches around services that require specific URL authentication inside Devin's machine.

For enterprise oversight, there is now a centralized admin console to manage multiple workspaces, access controls, and billing. Alongside its standard $500 monthly tier, which includes unlimited seats, Cognition quietly activated usage-based billing on January 9. Once a company burns through its monthly ACU allocation, developers can keep working past the cap with custom spending limits.

Companies like Ramp, OpenSea, Nu Bank, Lumos, and Curai Health are already testing Devin. But independent benchmarks show the software still has sharp edges. Researchers at Answer.ai tested Devin across 20 tasks and recorded 14 failures, 3 successes, and 3 inconclusive runs. They noted that Devin frequently wrote overly complex code and failed unpredictably, even on tasks similar to ones it had previously solved.

Why it matters

Devin 1.2 shows Cognition moving past the flashy demo phase and tackling developer logistics. Features like Slack voice notes and persistent browser cookies sound unglamorous, but they directly remove the friction that keeps engineering managers from assigning tickets to autonomous agents.

Still, at $500 a month plus overage fees, reliability remains the elephant in the room. If senior engineers spend more time untangling Devin’s over-abstracted code and verifying false positives than writing features themselves, the ROI collapses quickly. In our view, Devin is evolving into a capable junior pair-programmer for boilerplate chores and refactoring, but enterprises looking to replace human development loops will find the guardrails still need serious human supervision.