Our mission is to make agent work something you can rely on without reading all of it.
Agents now write the code, the decisions behind it, and the message saying it is finished. Each of those is a claim, and nothing makes a claim true. People cannot read everything agents produce, so trust has to come from checks a machine can run. How far agent work scales depends on how fast it can be verified.
Agents are usually described with two kinds of thinking: System 1, the fast guess, and System 2, the slow reasoning. Checking their work needs a third kind underneath both, one that does no guessing at all. We call it System 0.
Take the claim an agent makes most often: "this is done, and here is the test."
| Question | Answered by | Time | |
|---|---|---|---|
| System 0 | Does the test exist, and is it linked properly? | Deterministic checks: urzua | milliseconds |
| System 1 | Is this the right test? Does it test what the task says? | A probability from a decision model | about a second |
| System 2 | What does this test actually do? | Reasoning by an LLM | seconds to minutes |
Ask in that order. Each system hands the next only what it could not settle. System 0 goes first because its answer is the same on every run.
Paste this into your coding agent:
Read https://github.com/urzuahq/urzua/blob/main/docs/reference/install.md and https://github.com/urzuahq/urzua/blob/main/docs/reference/adopt.md, then set this repository up on urzua.
Or install it yourself:
tag=$(gh release view --repo urzuahq/urzua --json tagName --jq .tagName)
curl -fsSL --proto '=https' --tlsv1.2 "https://raw.githubusercontent.com/urzuahq/urzua/$tag/scripts/install.sh" | sh -s -- --version "$tag"| Project | What it checks | When |
|---|---|---|
| urzua | What the records say: this was decided, this is done, this is the evidence. A record engine configured entirely by the repository that uses it. Decision-model checks are on the way for the questions a rule cannot settle. | Every commit and pull request |
| decision-agent-kit | What the agent says in the session: it works, it is finished. Scored by a decision model. | The end of every turn |
- Evidence over assertion. "Done" has to point at something that exists.
- People own decisions. Agents draft; a named person accepts.
- Built for machines to read. Every result is data an agent or CI can act on.
- Honest about limits. A check says what it was measured to catch and what it misses.