← All speakers

Bio, Work & Ideas

Scott Wu

Conference affiliation: Cognition · 2025

On this page

Scott Wu is the co-founder and chief executive of Cognition, the company behind Devin, an autonomous software engineer designed to investigate codebases, execute development tasks, test changes, and collaborate with human developers. He is building toward a software industry in which engineers delegate substantial implementation work while retaining responsibility for architecture, product decisions, and judgment.

From competitive programming to Cognition

Raised in Louisiana, Wu won three consecutive International Olympiad in Informatics gold medals from 2012 through 2014, placing first with a perfect score in his final competition. He worked on performance engineering at Addepar, attended Harvard, and represented the university in the International Collegiate Programming Contest before leaving to build companies.

In 2017, Wu co-founded Lunchclub with Vlad Novakovski, developing an AI-powered professional-networking platform that matched people for meetings. After leaving in 2022, he co-founded Cognition with Steven Hao and Walden Yan in late 2023 to develop agents capable of navigating the practical complexities of software engineering.

Devin’s early demonstrations included building a mobile-friendly React game from conference-speaker data and implementing a search feature inside Cognition’s own product. The latter involved a prepared development environment, repository-specific instructions, a Git pull request, and revisions responding to human feedback—a concrete example of agents operating inside ordinary engineering workflows.

The ideas behind Devin

  • Autonomous software engineering: Wu treats programming as an iterative process spanning terminals, browsers, documentation, development servers, and code review. Useful agents must investigate problems, revise their plans, run software, verify results, and recognize when they need clarification.
  • Devin Playbooks: Repeatable procedures helped Devin tackle JavaScript-to-TypeScript conversions and Angular or Java upgrades before expanding into isolated bug fixes and multifile changes. Remote development environments and repository snapshots supply the infrastructure for running linters, continuous-integration checks, and tests.
  • DeepWiki: As agents take on more ambiguous work, they need structural knowledge of repositories, including call hierarchies, commit history, and relationships between files. DeepWiki makes that codebase understanding available to human developers, while Devin 2.0 integrates investigation, planning, and collaborative execution.
  • Parallel agent delegation: Wu argues that waiting for servers, logs, and tests limits synchronous coding assistance. His case for assigning multiple agents concurrent tasks envisions engineers guiding several independent workflows and intervening when judgment or additional context is needed.

In July 2025, Wu announced Cognition’s acquisition of Windsurf, bringing its developer environment together with Devin’s cloud-based workflow. By June 2026, Devin Desktop, introduced with Jeff Wang, combined Windsurf’s editor foundation with local and cloud agents, shared context, direct code editing, and support for compatible third-party agents.

In September 2026, Cognition reported a Series E above $2 billion at a $48 billion valuation. Its Fusion release for Desktop and CLI extends the product chronology: a frontier lead agent handles planning and review while delegating execution to a sidekick such as SWE-2.

Read the topics behind these talks

2 conference talks

Key ideas

Scroll to read ↓

As coding agents take on longer tasks, the hard problem shifts from following instructions to understanding repositories, clarifying intent, and testing their own work.

  • How much work can an agent do before you intervene?
    0:43 ↗
  • Repetitive migrations: follow the procedure, retain the corrections
    3:31 ↗
  • Small fixes still need a working repository
    6:22 ↗
  • Broader bugs require relationships, not just source text
    8:14 ↗
  • Explore together before handing off the task
    10:41 ↗
  • A backlog adds routing and escalation decisions
    12:51 ↗
  • Complete pull requests need an iterative testing loop
    14:04 ↗
  • The next doubling asks a different engineering question
    14:34 ↗

Key ideas

Scroll to read ↓

A name-game website, a production search bar, and a shared development machine show how coding agents turn instructions into software—and where human judgment still matters.

  • A name game built through feedback
    0:35 ↗
  • A search bar in Devin’s own repository
    3:45 ↗
  • From a hacker house to an agent
    5:36 ↗
  • Why code gives agents a useful feedback loop
    7:43 ↗
  • Supervision needs more than a chat window
    9:53 ↗
  • Delegating work asynchronously
    11:52 ↗
  • Deciding what to build remains engineering work
    13:16 ↗
  • Cheaper implementation could mean more software
    15:07 ↗
  • Access and executable verification
    16:27 ↗
  • How do junior engineers learn?
    17:34 ↗
  • Progress has to come across the stack
    18:42 ↗

References