← All speakers

Bio, Work & Ideas

Steve Yegge

Conference affiliation: Icon · Gas Town · 2026

Steve Yegge is a software engineer, technical writer, and independent AI adviser who created Gas Town, an open-source system for coordinating autonomous coding agents, and Beads, its persistent work-tracking foundation. His career spans developer infrastructure at Amazon and Google, engineering leadership at Grab and Sourcegraph, and a growing campaign to make AI-generated software more reliable, economical, and secure.

Yegge began by writing 8086 assembly at Geoworks, then worked at Amazon from 1998 to 2005 and Google from 2005 to 2017. At Google, he created Grok, an internal code knowledge graph for navigating large software systems. His widely circulated Platforms Rant contrasted Amazon’s API-oriented organizational discipline with Google’s platform strategy, establishing him as an influential, irreverent commentator on engineering culture and software architecture.

He subsequently led advertising and data-monetization engineering at Grab from 2017 to 2020. After returning to his long-running game project, Wyvern, he joined Sourcegraph in 2022 as head of engineering, worked on its Cody coding assistant, and later moved into an AI-focused individual-contributor role. His professional biography places his Sourcegraph tenure through 2025; he now works independently on agent orchestration and advises engineering organizations.

With Gene Kim, Yegge coauthored Vibe Coding, extending his interest in developer tools into a vision of engineers specifying outcomes, dividing work, and supervising systems that generate software.

The technical convictions behind his work

  • Orchestration beats a single oversized agent. Yegge advocates specialized agents for planning, implementation, review, testing, and integration. His vision for AI-native development treats limited context windows, human coordination costs, and indiscriminate use of expensive models as systems-design problems.
  • Durable work tracking makes autonomy practical. Released in October 2025, Beads preserves tasks and context across agent sessions. Gas Town, open-sourced in January 2026, coordinates parallel agents, supervisory roles, and merging around that shared ledger. Related initiatives include Gas City and the federated Wasteland work board.
  • Security belongs inside the generation loop. Yegge recommends separate review passes for security, correctness, performance, and coding standards; direct agent access to security scanners; least-privilege credentials; and supervisory agents. His agent-security analysis highlights malicious hallucinated dependencies, prompt injection, and a scan identifying 241 vulnerabilities in his own game after AI-assisted hardening.
  • Software survives by saving agents computation. In Software Survival 3.0, he argues that databases, search systems, and specialized tools remain valuable when they encode expertise and reduce expensive model inference.

For Yegge, production readiness ultimately depends on behavior, not spectacle: he has argued that carefulness matters more than headline capability when AI systems touch real customers.

In his August 2026 essay The Shape of Things to Come, Yegge describes a personal change in direction: his Gas Town workflow fell apart, and he gave up building reusable harnesses in favor of Wheelhouse, a private, closed-source orchestrator for Wyvern. His account describes his own workflow and approach to reusable harnesses.

Read the topics behind these talks

2 conference talks

Key ideas

Scroll to read ↓

Coding agents resemble powerful drills without the machinery that makes them precise. Steve Yegge and Gene Kim explore what coordinated agents, conversational development, and organizational change might put in their place.

  • A coding agent is an electric saw, not yet a CNC machine
    1:27 ↗
  • The adoption gap comes before the automation revolution
    2:52 ↗
  • One giant ant, one oxygen tank, and a swarm of specialists
    4:42 ↗
  • A deliberately provocative farewell to the familiar IDE
    6:34 ↗
  • Gene Kim: changing software production changes the organization
    7:43 ↗
  • From annual releases to programming through conversation
    11:09 ↗
  • FAAFO: faster, ambitious, autonomous, fun, and optionality
    14:05 ↗
  • A small delivery team, a Neo4j dashboard, and leaders who ship
    17:22 ↗
  • Trust is predictability; experience helps, but coordination remains hard
    21:04 ↗

Key ideas

Scroll to read ↓

Steve Yegge traces the risks from hallucinated dependencies to overprivileged agents, and explains why security tools, repeated review and explicit ownership must accompany autonomous work.

  • What happens when code ships ten times faster?
    0:41 ↗
  • A dependency can work perfectly and still be a backdoor
    2:55 ↗
  • Put findings where the code is being written
    4:49 ↗
  • A reassuring hardening pass still missed findings
    7:22 ↗
  • Let agents offload security work to tools
    9:33 ↗
  • Security belongs at the beginning and the end
    11:47 ↗
  • The attack surface includes your family
    13:34 ↗
  • Why the capability curve can look flat
    15:30 ↗
  • Gas Town: the missing tasks were already implemented
    17:22 ↗
  • Who watches an agent that can take action?
    19:02 ↗
  • Prompt injection needs an owner
    20:46 ↗

References