Why We Built Switchyard
AI coding tools have gotten remarkably good. In the past year, we’ve watched single-agent assistants go from suggesting autocomplete to writing entire features. But as we pushed them harder — on real codebases, with real complexity — a pattern emerged.
The single-agent ceiling
Ask an AI agent to “add OAuth with GitHub” and you’ll get… something. Maybe it works. Maybe it half-works. Maybe it edits the wrong files or forgets to write tests or introduces a security hole that nobody catches because there’s no review process.
The problem isn’t intelligence. The models are smart enough. The problem is structure.
When a human team tackles a complex feature, they don’t hand it to one person and hope for the best. They break it down. They assign specialists. They write tests first. They review before merging. There’s a process — a pipeline.
AI coding tools have none of this.
What we wanted
We wanted a system that works the way good engineering teams work:
- Decompose complex tasks into discrete, scoped subtasks
- Assign each subtask to a specialized agent (coder, tester, reviewer, security analyst)
- Run those agents in parallel on isolated git worktrees — no file conflicts
- Review every change through a structured pipeline before anything merges
And critically: keep all source code on our own infrastructure. No shipping proprietary code to third-party servers. No hoping someone else’s security is good enough.
The architecture
Switchyard splits into two parts:
The control plane runs on Cloudflare Workers. It manages task orchestration, pipeline state, analytics, and the dashboard. It never sees your source code — only metadata like task status, step counts, and token usage.
The relay runs on your infrastructure. A single Docker container that manages git repos, spawns agent containers, stores execution traces, and keeps your API keys encrypted locally. It connects to the control plane via WebSocket to receive orchestration commands and report metadata.
The result: all the convenience of a managed SaaS platform, with none of the security trade-offs.
What’s next
Switchyard is now in early access. The core pipeline is working end-to-end — task decomposition, multi-agent execution, human review, and merge. We’re looking for engineering teams who want to be among the first to orchestrate AI agents on their own terms.
If that sounds like you, get started or read the docs.