The Future of Coding Agents: From Workers to Factories
Development History
- v1: vibecoder (TypeScript/Temporal) — too cumbersome for micro-workflows
- v2: vc (Go) — monolithic, focused on making agents better
- v3: Python Gas Town (PGT) — quantity over quality, “agent village” concept
- v4: Gas Town (Go) — complete rewrite, achieved liftoff Dec 2025
Why Go for AI Orchestration
- Token efficiency: TypeScript wastes tokens on complex type manipulations
- Simplicity: Go’s boring nature makes AI-generated diffs easier to verify
- Portability: Native binaries easier to distribute
Key Predictions
- Shift from “super-worker” agents to “agent colonies” / factories
- Orchestrator API surface becomes competitive differentiator
- Factory farming code: success from orchestrators turning pet agents into cattle
- Big companies will struggle with velocity of AI-driven development
- By Summer 2026, OSS models catch up to late-2025 frontier models
The “Desire Paths” Approach to Agent UX
- Tell agent what you want
- Watch what it tries to do (even if it fails)
- Implement the tools/structures the agent expected to find
- Repeat until tool works exactly how agent believes it should