Descriptions:
Walden Yan, co-founder and CPO of Cognition (the company behind Devin), and Cole Murray, creator of Open Inspect, join the Latent Space podcast to share firsthand data on how autonomous coding agents have transformed software development in early 2026. The episode centers on what they call the “background agent” paradigm — a shift triggered by capability jumps in models including Claude Sonnet 3.7, Opus 4.5, and GPT 5.2, which enabled agents to go from a specification to a completed pull request with minimal human intervention.
The internal metrics are striking: Devin’s merged PR count grew 7x over roughly two to three months, while Cognition’s engineering headcount grew only about 10%. Devin’s own repositories went from 16% AI-generated commits in January 2026 to 80% in March. The conversation digs into the architectural decisions behind making this work at scale — including why off-the-shelf MCP integrations frequently fall short for production workflows (Slack being a concrete example requiring custom webhook handling), why memory and knowledge retrieval remain an unsolved problem that the team has deliberately avoided rushing, and how the distinction between “computer use” and genuine test reasoning shapes how Devin approaches frontend-backend integration testing.
Yan also explains the origin of the term “context engineering” as an upgrade from “prompt engineering,” and why the team’s internal guidance discourages multi-agent systems as a default architecture. Cole Murray shares lessons from building Open Inspect, including where MCP’s simplicity promise breaks down when the spec gets complex.
📺 Source: Latent Space · Published May 28, 2026
🏷️ Format: Interview
