Switchyard
AI work with independent review and a record of what happened.
Models differ in capability and availability. I wanted a consistent workflow that could carry a task through independent review to a checked result.
Launch terminal demo — Sped-up demonstration with fictional project data. No model requests are sent.
Explore the process
- A task with a checkable result
The brief states the work, the relevant context, and how to verify the result. The kind of task and its data constraints determine which providers are eligible.
- Choose an eligible model
Switchyard selects a provider for the task. If that provider is unavailable, a defined fallback path chooses another eligible provider without dropping the task’s constraints.
- Keep the author out of its review
Separate model families review the change. Findings return to the development process for assessment, fixes, and another check.
- Check what was actually produced
Artifacts and test results are evidence. A model saying it finished is not enough. The record retains the result and the routing and review history.
Beyond code
The same system routes research and inspection tasks. Each task keeps its own provider eligibility rules, so a research request can use a different set of models from a code change.
Inside the demo
A fictional photo-import feature moves through authoring, review, fixes, and verification. Claude Code coordinates the session, with Grok Build and OpenCode–GLM shown in the workflow.
Scroll up to pause and inspect the terminal. Return to the bottom to continue. The TM2 tab contains another small project.
Example result
This is a synthetic illustration, not a record of a production incident.
A first eligible provider did not return a usable artifact. A permitted fallback ran under the same data restrictions. Independent reviewers required a correction. The accepted result records the checks that passed.