The Equation Series
Menu

Part III: Building What Works

Chapter 16

Agents and Work

On 12 March 2024 Cognition, a San Francisco start-up of about ten people founded the previous year by the competitive programmers Scott Wu, Steven Hao and Walden Yan, introduced Devin as "the first AI software engineer" (company). Devin worked in its own sandbox with a shell, an editor and a browser. Given a task, it planned, wrote and ran code and reported back. Cognition said it resolved 13.86% of a subset of SWE-bench, a benchmark of real GitHub issues checked by tests, end to end, against about 1.96% for the previous best unassisted system (company, March 2024). Its videos showed Devin setting up environments, fixing bugs and completing a freelance job posted on Upwork. There was no public product, only a waitlist. Within weeks Founders Fund led a $175 million round at a $2 billion valuation (April 2024, reported), on about $21 million raised before.

The backlash came in April. A software developer's video, "Debunking Devin", argued that the Upwork demonstration misrepresented the job: the agent fixed errors in code it had itself introduced and did not deliver what the client had asked for (reported). Early users reported that Devin often got stuck or produced wrong results, and an evaluation by Answer.AI researchers in January 2025 reported that it succeeded on 3 of 20 tasks (reported, verify). Cognition kept going. Devin became generally available at $500 a month in December 2024, a $20 entry plan followed with Devin 2.0 in April 2025, and annualised revenue rose from about $1 million in September 2024 to about $73 million in June 2025 (company figures, reported). By May 2026, after buying the rest of the Windsurf editor, Cognition reported combined annualised revenue of about $492 million and raised more than $1 billion at a $26 billion post-money valuation (reported).

This chapter is about agents, models that act rather than answer, and the claim that they let software companies sell into the labour budget rather than the IT budget. It asks what an agent can complete without a person, what a completed task costs, and what happens to the firms whose work it takes.

The preview ends here. The full chapter is in the book.

In this chapter

  1. 16.1Why long tasks are hard
  2. 16.2From demonstration to product
  3. 16.3Services as software
  4. 16.4The integrators and the sell-off
  5. 16.5A worked example: cost per merged task
  6. 16.6How an expert reads an agent business