Switchyard
Coding agents are fast. Trusting what they ship is the slow part. Switchyard makes them prove it.
InfrastructureRepository
The Problem
Running several agents at once turns into babysitting: which one is stuck, what is it costing, and did the tests actually run? Switchyard puts the whole fleet on one dashboard, caps spend, and refuses to call a task done until the proof is attached.
The Solution
A local control plane for coding agents. It claims tickets, runs workers in isolated worktrees, opens pull requests, and holds every change behind eight evidence-backed gates, so a person only has to press merge.
What it does
- Dispatcher claims tickets within worker and daily spend limits, one isolated worktree per agent
- Eight gates from implemented to verified in production, each carrying its own evidence
- Live crew view with every worker's in-flight tool calls
- Reviewer edits become rules the next worker reads before it starts
- Slack push the moment a task needs a human
What took work
- Keeping a live dashboard honest from an append-only event journal instead of agent self-reports
- Enforcing a hard spend cap across concurrent workers that cannot see each other
- Running TypeScript source directly with a V8 compile cache so the daemons start fast with no build step