Devin’s 80% Moment: Background Agents, 7x PRs, & End of Hand-Held Coding — Walden Yan & Cole Murray
Summary
Background coding agents became commercially different around December 2025, when Opus 4.5 and GPT-5.2 could move from a good specification to a finished pull request with “very little friction.” Cognition’s internal evidence is unusually concrete: Devin’s commit percentage across Devin repositories rose from 16% in January to 80% in March, while merged-PR usage grew roughly 7x over two to three months against only about 10% engineering-headcount growth. Cole said Sonnet 3.7 had already enabled Cognition to strip out agent machinery that greater model intelligence made unnecessary.
The standalone agent-orchestration layer looks economically thin unless a vendor owns infrastructure, distribution, or the enterprise adoption journey. Cole open-sourced OpenInspect because background agents will be critical company infrastructure, but declined to compete for “$20 a seat” while model and sandbox providers capture much of the value: “What are you actually selling?” Cognition’s answer is a combined agent, compute stack, integrations, and engineering team that helps enterprises reorganize work around AI.
Security and portability favor Cognition’s architecture of “separating the brain from the machine,” despite the additional state-management complexity. An in-box harness is simpler, but secrets and unpredictable agent behavior coexist inside one sandbox; an out-of-box brain can keep sensitive control logic separate while each machine receives only the credentials defining that user’s permitted scope. The architecture also supports existing dev boxes, full VMs, OS-specific environments, and Android development in beta.
The defensible engineering is increasingly below the chat loop: repo setup, fast resumable machines, credentials, and realistic local applications. Teams still rely on “go talk to Bob and get the secrets,” while raw EC2 machines once left Devin “out cold for like 10 minutes.” Cognition built storage that restores work proportional to the filesystem diff rather than an entire terabyte-scale disk, discovered that network-backed filesystems made every grep a network operation, and owns enough infrastructure to work toward VPC, on-prem, and GovCloud deployments.
Application testing is a reasoning-and-orchestration problem, not merely a computer-use problem. Clicking is “emit the right coordinates”; testing a frontend-and-backend change means launching compatible versions, satisfying admin or feature-flag conditions, coordinating sessions, and discovering the exact action that triggers behavior. The discussion says some cases require orchestrating multiple frontier models because no single model can complete the entire chain, while labeled videos and screenshots compress the human verification loop.
Enterprise value depends on bidirectional workflow integration, where generic MCP connections often stop short. A useful Slack coworker must receive webhooks, respond naturally, preserve shared sessions, and avoid spamming threads; a useful GitHub agent must resolve comments and conflicts without looping on its own reviewer. The discussion calls for something “more expressive than MCP” that goes both ways, while Swyx notes that owning a critical integration may make sense when nearly every agent session traverses it.
Memory and multi-agent systems remain promising but operationally immature. Cognition’s memory system must infer durable preferences without universalizing one-off requests, retrieve among thousands of memories without flooding context, and survive model changes; the team is exploring filesystem-like memory that agents navigate themselves. Multi-agent work similarly succeeds most reliably as one manager decomposing isolated tasks, with subagents acting like context-compressing tool calls—not free-form swarms—although agents’ newfound ability to reject bad instructions makes real collaboration more plausible.
Near-term returns concentrate in auto-triage, security review, support investigation, and letting non-engineers initiate bounded code changes—but governance remains mandatory. Walden’s warning is that “your codebase regresses to your worst engineer” when unchecked AI-generated patterns become future training context; Cognition’s no-review experiment lasted about two weeks before duplication made simple changes painful. Swyx cited spending of roughly $1,000–$5,000 per engineer, and the response acknowledged that figures can go much higher; hybrid frontier/subfrontier routing is expected to become a major cost lever.
Deep dive
Not yet available upstream; scheduled sync will retry.