Gas Town, Beads, and the Rise of Agentic Development with Steve Yegge

What It Covers

Steve Yegge discusses the evolution from chat-based AI coding to full agent orchestration, introducing Beads (a Git-backed task graph system) and Gastown (a multi-agent orchestration framework). He explains how developers now manage fleets of coding agents, the shift from writing code to orchestrating work, and predictions for how AI transforms software development workflows and team structures.

Key Questions Answered

• •Beads Task Graph System: Beads combines three components to create shared memory for AI agents: a directed graph structure for task relationships, SQL database queries for agent access, and Git ledger for complete work history. This enables agents to track work surfaces, maintain forensics through "discovered from" fields, and never lose context. The system transitions to Dolt database to eliminate merge conflicts and enable field-level resolution instead of issue-level conflicts.
• •Context Window Strategy: Developers split between two approaches - minimizers use small, ephemeral tasks with tools like Polecats for well-specified throwaway work to optimize cost and performance, while maximizers load heavy context for strategic design decisions using crew workflows. Successful orchestration requires switching between both modes: high-context conversations for difficult architectural corners and low-context decomposition for factory-style code production across parallel agents.
• •Landing the Plane Protocol: AI agents exhibit overconfidence, declaring completion prematurely with celebratory messages despite incomplete work. The "land the plane" prompt leverages their affinity for bureaucracy and checklists, forcing systematic verification even when context windows near exhaustion. This protocol transforms unreliable completion into consistent task closure by exploiting their OCD-like tendencies to check off acceptance criteria, preventing forgotten steps and compassion failures during handoffs.
• •Trust-Patience Inverse Relationship: As developers accumulate hundreds to thousands of hours practicing with AI coding agents, trust in predictable outcomes increases while patience for sequential work decreases. This drives adoption of parallel agent workflows - developers who confidently predict agent success on repetitive tasks immediately spawn additional agents rather than waiting. The gateway drug moment occurs when spinning up a second agent while the first handles known-good work patterns.
• •Specification Through Iteration: Gastown operates on first-approximation philosophy rather than heavy upfront specification, except for critical infrastructure like database migrations. Developers spend time in two modes: minimizing context for well-specified ephemeral tasks or maximizing context for design work requiring rich understanding of why, not just what. The shift moves engineers toward heavy thinking work while agents handle coding, resulting in frequent naps from mental exhaustion managing factory-like agent teams.
• •Thermodynamic Survival Principle: Software products survive AI disruption by saving tokens through computational efficiency. Tools like Serena (LSP servers), databases, calculators, and infrastructure systems that perform math cheaper than LLM matrix multiplications become preferred by agents seeking lowest energy states. Products must overcome activation energy barriers through direct model training partnerships with OpenAI, Anthropic, and Google, plus become systems of record with exclusive data access to justify token expenditure.