Cognition's Scott Wu launches agent-native IDE for Devon and says AI will beat IMO gold this year

Summary

Scott Wu, CEO of Cognition, launched an agent-native IDE for Devin on April 3rd — a full development environment designed around asynchronous, parallel workflows rather than the synchronous autocomplete model that most AI coding tools are built on. The core idea is that engineers manage a team of Devin agents simultaneously, delegating tasks and reviewing progress rather than sitting in a back-and-forth loop with a single assistant. The release includes a collaborative planning experience at the start of each session, plus features called Devin Search and Devin Wiki. Entry-level access starts at $20 and is fully generally available at app.devin.ai.

Wu frames the product around a distinction between two markets. The first is consumer and hobbyist coding — building websites, games, and personal tools — which he says is already taking off and will keep expanding as AI lowers the barrier to entry. The second is enterprise software engineering, where he expects AI to deliver roughly a 5x productivity gain for professional engineers rather than replace them outright. In his view, the complexity of real-world codebases — legacy systems, COBOL, idiosyncratic environments, millions of lines of code — means deep technical work still requires engineers who can reason about architecture. The bottleneck isn't capability in isolation; it's getting AI to work reliably across all of that messiness.

On the bubble question, Wu is unambiguous: AI value is underrepresented, not overstated. He draws a parallel to Sam Altman's 2015 "Bubble Talk" post, which predicted that the top unicorns of the day — Uber, Airbnb, Stripe, SpaceX — would be worth more than $200 billion by 2020. Several of them now exceed that figure individually. Wu argues AI companies have at least as strong a fundamental case, because the addressable value isn't a single workflow made 1% easier — it's multiplying output across every knowledge worker.

On multimodality, Wu argues that visual and interactive feedback loops are essential for real-world coding tasks. Dumping a million-line codebase into a context window and asking for a fix doesn't work for humans or AI. What works is iterating — running a front end locally, clicking around, reading logs — and multimodal capability is what gives AI agents that same incremental, evidence-based approach to problem-solving.

The sharpest claim Wu makes is about competitive programming benchmarks. Gennady Korotkevich — widely regarded as the greatest competitive programmer of all time and, Wu confirms, an employee at Cognition — has an internal bet running on whether AI will surpass him this year. Wu says most people at the company think it probably happens in 2025, and if not, in 2026. AI took silver at the IMO last year; Wu says he'd be surprised if it doesn't win gold this year. Polymarket currently prices that outcome at 59%.