Steve Yegge

Yegge completed his University of Washington computer-science degree in 1995 while working at Geoworks. At Amazon he worked on a migration toward services and led developer-productivity teams. At Google he created Grok, the code knowledge graph later known as Kythe. He subsequently led engineering groups at Grab. At Sourcegraph he moved from head of engineering to coding on Cody, before leaving in October 2025 for independent work.[1]

Work that survives the next session

Beads stores tasks, notes and work relationships in a version-controlled ledger that agents can query and update. A replacement session recovers task state from the ledger. Gas Town adds orchestration around that shared memory. Workers execute tasks, a witness monitors them, a refinery serializes merges and a mayor coordinates work across groups. The design treats integration and recovery as explicit responsibilities. Yegge’s project guide also distinguishes the standalone ledger from the larger orchestrator, allowing developers to adopt persistence without deploying the whole system.[2]

Writing about the changing craft

Yegge has written about programming languages, engineering productivity and software culture for decades. With Gene Kim he co-authored Vibe Coding, a book about building production software with generative AI, chat interfaces and agents. His current tools make that interest concrete: they address what happens after a useful generation, when work must be remembered, coordinated, reviewed and integrated into a maintained codebase.[4]

In “Fences, Not Sandboxes,” Yegge describes operating a collection of agents to develop his long-running game Wyvern. He argues for expressing organizational decisions as explicit rules, with mechanisms that enforce those rules and make exceptions visible. His account includes failures, such as an unplanned release, alongside the gains from delegating work. The essay’s central proposal is that sustained autonomy requires a record of decisions and enforceable boundaries around actions. It is a practitioner’s argument drawn from his own system, rather than a general proof that agents can safely govern arbitrary software organizations.[3]