Orchestrate
A control plane that makes AI coding agents prove their work.
One prompt in. The orchestrator plans, asks for what it needs, then runs workers in parallel and collects their reports for review.
Overview
Coding agents, held to evidence.
I love working with AI coding agents, and the more work I hand them, the more their reports need checking. Orchestrate coordinates Codex, Claude, and other agents under one review surface. It plans the task, runs workers in isolated git worktrees, and captures screenshots, DOM snapshots, console output, network traces, and assertions, so every claim can be verified and replayed.
I lead development of Orchestrate. It is open source under the MIT license.
Why I built it
I love working with agents. I wanted to trust them.
Coding agents let me move fast on hard problems, and I work with them by choice. The catch is how they report back: in prose. An agent says the bug is fixed and everything works, and you either believe it or redo the check yourself.
Orchestrate changes the question from what the agent says to what it did. Every run leaves evidence behind, so reviewing an agent means checking that evidence, not redoing the work.
How a run works
One prompt, six recorded steps.
Intake
The operator submits a prompt from the web app, the desktop app, or by voice.
Plan
The orchestrator drafts a scoped plan and an acceptance contract before any code is written.
Worker
A provider-neutral worker, Codex or Claude, runs in an isolated git worktree.
Evidence
Screenshots, DOM snapshots, console output, network traces, and assertions are captured automatically.
Review
A reviewer gate checks the structured acceptance criteria, and the operator approves or redirects.
Ledger
Every transition is appended to an event-sourced log, so the run can be replayed and audited.
Results
Measured on VerifBench.
VerifBench, the benchmark that accompanies Orchestrate, measures verification debt across 8 harnesses, 12 tasks, and 6 metrics.
- VerifBench score
- 98.75%
- Claude worker, 711 of 720
- Evidence completeness
- 10/10
- Cohort average 7.4
- Same-backend lead
- +7.08 pts
- Opus 4.6, against the strongest existing surface
- Harnesses evaluated
- 8
- 96 task runs, 576 metric scores
What this sample shows
Building with agents, and for them.
Multi-agent systems, AI coding workflows, automatic evidence capture, event-sourced architecture, and a product built around trust in AI output. This is one sample of that work, not the full list.