Building Tatou Around Evidence

Tatou started from a simple frustration: coding agents can sound confident even when their work is not proven. A model can say it edited a file, ran a test, or fixed an issue, but the useful question is whether there is evidence behind that claim.

The product direction is an execution layer for AI coding work. A user gives a goal, a manager turns it into tasks, a builder proposes changes, tools execute those changes, checks run, and a reviewer compares the result against the evidence. The model is part of the workflow, but it is not the source of truth.

The important design choice is traceability. Tatou records model calls, tool actions, verification commands, changed files, and user-facing claims with OpenTelemetry. Claims like "tests passed" or "file saved" should point to command output, file evidence, or GitHub state. If they cannot, they should be treated as unsupported.

That same trace data also feeds the improvement loop. When Tatou notices unsupported or refuted claims, it can create an improvement candidate, deduplicate it against existing issues, and route it back through the normal spec, implementation, test, and review process.

I am still shaping the product, but the principle has stayed stable: trustworthy AI work needs evidence, not just good wording. Tatou is my attempt to make that principle feel natural in everyday coding workflows.