Steve Yegge
Conference affiliation: Icon · Gas Town · 2026
Related work and projects
- Welcome to Gas Town ↗steve-yegge.medium.com · Source ↗
- The Shape of Things to Come: The Continuous Thunderdome ↗
Yegge’s August 2026 account of his move to private Wheelhouse and Wyvern.
yegge.ai
Steve Yegge is a software engineer, writer, and independent AI adviser who created Beads and Gas Town, open-source tools that preserve coding agents’ work and coordinate their sessions. Earlier, he built Google’s internal code knowledge graph, Grok, and wrote the Platforms Rant. His recent work puts his enthusiasm for autonomous development through the demands of a live multiplayer game: agents must carry assignments across sessions, integrate their changes, and serve players who may not welcome every increase in development speed.
On this page
Based in Kirkland, Washington, Steve advises engineering organizations adopting AI. His essays combine technical argument, organizational criticism, and humor, drawing on work inside Amazon, Google, Grab, and Sourcegraph.
From operating systems to code intelligence
Steve began at Geoworks, writing 8086 assembly for desktop and mobile operating systems. His responsibilities also included engineering project management and delivery of HP’s OmniGo handheld devices. At Amazon, from 1998 to 2005, he led teams working on customer-service applications and developer tools. His writing began in the company’s internal developers’ journal and grew into Stevey’s Blog Rants.
At Google, Steve created Grok, an internal code knowledge graph used with Code Search to help developers navigate, understand, and refactor the company’s codebase. This was infrastructure for understanding relationships within a large software system, beyond finding matching text. His accidentally public 2011 Platforms Rant contrasted Amazon’s API-oriented engineering discipline with Google’s platform strategy, arguing for systems whose capabilities other developers could build upon.
After Google, Steve led advertising and data-monetization engineering at Grab, then returned to Wyvern, his long-running online game. He joined Sourcegraph as head of engineering in September 2022, extending his work on code intelligence beyond Google’s internal infrastructure. He worked on its Cody coding assistant and moved into an AI-focused individual-contributor role in February 2024. His public career history records his departure in October 2025 and subsequent independent work on Gas Town and Beads.
With Gene Kim, Steve coauthored Vibe Coding, published in 2025. The collaboration brought his experience with coding assistants into a broader account of how AI could change software development and engineering organizations.
Giving agents work they can continue
Steve’s agent tooling addresses a practical mismatch: a coding session can end while the project’s obligations remain. Decisions, dependencies, unfinished assignments, and integration work need to survive the agent that first encountered them.
- Persistent work tracking. Beads stores issues and dependencies in a structured graph. Agents can find unblocked tasks, claim work, and attach findings to the relevant issue. Subsequent sessions can recover those records instead of relying on a person to reconstruct what happened. Steve originated the project; he credits the Dolt and Gas City teams as co-maintainers.
- Coordination across sessions. Gas Town combines the Beads ledger with persistent assignments, agent identities, messaging, supervision, and merge processing. Worker sessions can end while their assignments and history survive. Coordinators distribute tasks, supervisors watch for stalled workers, and a merge processor checks and integrates completed changes. Wasteland extends this work exchange across Gas Town workspaces.
- Interfaces shaped by agent expectations. In Software Survival 3.0, Steve describes adapting Beads to commands agents repeatedly expected to exist. He implemented those operations rather than continually correcting the agents through prompting. This “desire paths” approach treats retries and mistaken commands as evidence of where the interface should change.
The same essay develops his argument that software can retain value by saving agents computation. Git packages accumulated knowledge about managing changes; tools such as grep perform searches without spending model inference on the task. Steve’s framework weighs those savings against the cost of discovering and operating a tool. He presents it as a mental model, not a demonstrated law: useful software can still be passed over if agents do not know it exists or repeatedly struggle to use it.
Security inside the generation loop
Steve’s advocacy for faster development includes an uncomfortable consequence: more generated code can mean more vulnerabilities to investigate. In his agent-security discussion, he reported that Snyk found 241 vulnerabilities in his game after an AI-assisted hardening pass. The model’s confident completion did not establish that the work was secure.
He recommends separate review passes for correctness, performance, coding standards, and security, with scanners available directly to the agent doing the work. Findings should arrive while the agent can act on them. His warning about hallucinated dependencies explains another failure mode: attackers can publish malicious packages under names models mistakenly recommend, so a successful build and passing tests do not establish a trustworthy dependency.
For unattended agents, Steve advocates supervision and credentials limited to the actions they need. He also identifies prompt injection as a serious problem without claiming to have solved it. In a July 2026 post, he made careful behavior his overriding criterion for models handling customer-facing production work. That is his stated operating preference, rather than a comparative benchmark.
When the orchestration became the project
Wyvern, begun in 1996 and launched in 2001, became the setting for a consequential reversal. In his August 2026 essay The Shape of Things to Come, Steve described abandoning his own Gas Town setup after agents repeatedly tinkered with the orchestration instead of progressing on application work. He built Wheelhouse, a private, closed-source orchestrator tailored to the game, and came to argue that harnesses should be closely integrated with their applications.
Wheelhouse nevertheless returned to much of Gas Town’s shape: Beads, planning agents, implementation workers, separate reviewers, messaging, and merge processing. Steve’s reversal concerned reusable harnesses, while task tracking and specialized roles remained central to his approach. He also acknowledged the cost: work on Wheelhouse itself occupied roughly 20–25% of his Wyvern development effort at the time.
The game exposed a limit beyond coding. Steve says players asked for a roadmap because the pace of change made them anxious. He slowed feature launches and concentrated on quality. His agents could produce changes faster than his players wanted to absorb them.
Read the topics behind these talks
2 conference talks
▶ Watch ↗AI Engineer World's Fair 202622:32
Agentic Security: Permissions, Provenance, and the Agent Supply Chain
Explore the talk →Loading more talks and highlights…
References
- ai.engineer/talks/2026-the-year-the-ide-died ↗
- ai.engineer/talks/agentic-security-permissions-provenance-and-the-agent-supply-chain ↗
- steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04 ↗
- yegge.ai/gastown ↗
- yegge.ai/essays/the-shape-of-things-to-come/ ↗
- ai.engineer/worldsfair/2026 ↗
- ai.engineer/code/2025 ↗
