Steve Yegge

Steve Yegge is a software engineer and writer, coauthor with Gene Kim of Vibe Coding (IT Revolution, 2025), and the builder of Gas Town, an open-source toolkit for orchestrating AI coding agents. His contributions range from Grok, Google's internal code knowledge graph, to engineering leadership at Sourcegraph and the Beads work ledger for agent memory. His public views also changed substantially: skeptical of working agents in December 2023, he was advocating supervised agent clusters and fleets by March 2025.[5][6][1][4]

Yegge dates Beads' release to October 2025 and says he open-sourced Gas Town on January 1, 2026, after building it to run his own coding-agent work. Beads supplies the durable ledger beneath Gas Town: tasks, fixes, merge requests, and agents' notes become queryable work records that agents read and write and humans audit. A successor session can read those records after a crash or handoff and continue the task.[6]

Beads organizes work as a dependency-aware graph rather than a collection of Markdown plans. Its documented implementation uses Dolt, a version-controlled SQL database, with dependency tracking and detection of tasks ready to start. Git integration is optional; the database supplies the persistent memory.[7]

Gas Town separates execution, oversight, and integration into named roles. Polecats are worker agents with persistent identities and temporary sessions; a witness monitors workers within a project; a mayor coordinates across projects; and a refinery processes merges. The documented merge queue runs verification gates and isolates failed merge requests. Git worktrees preserve agent work separately from the lifetime of a conversation.[6][8]

These agent systems came after Yegge's work on code understanding at Google and Sourcegraph. That earlier work supplies context for his AI coding explanations: before coordinating agents, he built infrastructure that helped tools navigate and interpret existing code.[3][6]

Yegge began with 8086 assembly work at Geoworks. At Amazon, where his biography places him from 1998 to 2005, he led teams and programs in Customer Service technology and Developer Tools. He subsequently worked at Google and later led engineering for Grab's Ads & Data Monetisation business in Southeast Asia. Alongside that career, he wrote Stevey's Blog Rants and the internal 2011 “Platforms Rant,” which accidentally became public.[5]

In his 2022 account of Grok, Yegge recalled beginning the project at Google in 2008 after engineers complained that they could not find their way around the codebase. Grok harvested intelligence from production compilers, aggregated it into a queryable knowledge graph, and exposed APIs to other tools. Google Code Search belonged to a separate engineering team: it found starting points for exploration, while Grok supplied the relationships needed to navigate onward.[3]

Beyang Liu used Google's code search as an intern and later recalled missing it more than any other developer tool. That experience helped motivate Liu and Quinn to start Sourcegraph. Liu initially knew Yegge through his essays; watching Yegge's talk about Grok connected the essayist to the system he had used.[1]

Yegge announced joining Sourcegraph as Head of Engineering on October 4, 2022. In his account, founders Quinn and Beyang approached him, but a coding assignment during the interview made the opportunity compelling. While working in the codebase, he discovered that existing backend engineers had created SCIP, the source code intelligence protocol, and were moving beyond LSIF. He regarded LSIF as too lossy for a broader code intelligence platform and saw SCIP's extensibility as a way to pursue work he had begun with Grok.[3]

Discussing Sourcegraph's Cody assistant in December 2023, Yegge compared retrieval-augmented generation to consulting an expert who knows a particular codebase, and fine-tuning to on-the-job training. His point was that training does not remove the need to look up code-specific information. He also clarified that Sourcegraph's claimed “data moat” was its preprocessing engine: the ability to dissect a codebase into fine-grained semantic units and serve useful context.[1]

Yegge described fitting retrieved code into a limited context window as a bin-packing problem. Ranking results was only part of it: a request might need a function signature or a small expression rather than an entire function. Smaller semantic units, summarization, and testing across model configurations could help make that context useful. Beyang Liu named their hybrid of formal code intelligence and learned models “Normsky”; Yegge emphasized the testing required to make the combination work.[1]

In December 2023, Yegge said, “I'll believe in agents when somebody shows me one that works.” He favored improving context sources and giving users more control over the context sent to models. He credited Dominic with an earlier “nonstop Cody” feature that allowed parallel model requests to modify a source file, but recalled judging it slightly ahead of its time; some of its diff-handling scaffolding was reused for inline assistance.[1]

By March 22, 2025, Yegge was advocating coding agents and reported running two in parallel. He defined vibe coding broadly as letting AI do the work, with a material distinction: “For production, you pay attention; for prototypes, you chill.” He still emphasized task decomposition, thoughtful problem selection, supervision, validation, and verification.[4]

In Revenge of the junior developer, Yegge proposed six overlapping waves: traditional programming, completions, chat, coding agents, agent clusters, and agent fleets. His forecast put human developers above AI supervisors managing groups of coding agents—for example, separate groups handling bugs, features, and an architectural migration. He predicted fleets by early 2026 and acknowledged Quinn Slack for contributing most of the ideas. These were his projections for software work, accompanied by a call for developers to learn to use agents rather than wait for them to prove superior at every task.[4]