Cognition's Scott Wu: enterprise usage of Devon doubled in 6 weeks, Windsurf evolves from code editor to English-as-source-of-truth interface

Summary

Cognition's enterprise usage more than doubled in the last six weeks, Scott Wu says, driven by rapid adoption of AI agents handling end-to-end tasks. The company announced updates to Devon, its autonomous AI software engineer. Automated web app testing with screen captures, faster VM startup times, smoother Slack integration, and visible intermediate progress on agent work all reduce friction in the user experience. Wu describes these as clearing bottlenecks rather than fundamental capability leaps.

The concrete constraints are no longer model speed but the systems around the agent. npm installs, dependency resolution, frontend pulls, and other infrastructure tasks are where traditional software engineering optimization applies. Algorithmic tricks, indexing, parallelization, and async execution matter, but most of the work is basic product building.

Wu dismisses the broader narrative of a SaaS apocalypse. Software is one of the most deflationary industries ever, yet the world's largest companies remain software companies. The difference lies between price drops caused by collapsed demand and price drops from superior supply. The latter creates consumer surplus and drives growth. Devon's customers—banks, health insurers, private equity firms—use it for database migrations and modernization, then immediately ask how to accelerate the rest of their roadmap. More software needs to be built.

On AI progress, Wu sees the exponential curve continuing in a different form. About four to five months ago, Cognition stopped typing code. Almost none of the code they check into GitHub is written by humans anymore. As code generation became less of a bottleneck, new ones appeared: understanding codebases, review, and testing. Solving each requires good product design and model capability, not just better AI at writing from prompts.

Windsurf, Cognition's code editor, will evolve toward what Wu calls English as the source of truth. Over one to two years, the interface will shift from code files to specs, design docs, and diagrams. Developers will review decisions and product behavior rather than hunt bugs in code. Agent work happens in the background while human interaction moves to a higher layer of abstraction. Reading code matters less; reviewing what the system actually does matters more.

Wu points to recent work by Alex Lupaska and OpenAI researchers using language models to discover key lemmas and theorems in physics as a sign of coming scientific breakthroughs. He expects AI-powered actual scientific discovery to accelerate throughout 2026. Average consumers won't immediately feel the impact, but long-term gains in medicines, biology, and material science flow from the same capability. Material science particularly interests him as tangible progress: stronger, cheaper, lighter carbon fiber at Model 3 prices.