Evidence and checks for AI agents
Use your AI.
Check its work.
Grainulator works with Claude Code and Codex. It keeps the findings and sources your agent records together, flags gaps in evidence, and helps you check results before sharing.
See how the check worksSee what a check catches
Good yesterday.
Still good today?
A report is checked and ready. Later, a supporting test fails. Grainulator spots the change and stops the old report from being shared.
Follow the evidence
A real Grainulator check using sample data. Choose a step to see what changed.
Checking the report…
This example uses a saved result. It makes no new AI request.
—
See the exact evidence record
- Checked version
- —
- Current version
- —
How it works
Keep findings and sources together.
Grainulator stores the findings and source references your agent records, flags gaps, and helps your agent address failed checks.
- 01
Start with your task
Give your agent the question, background, and goal.
- 02
Check what matters
Choose a test or review step. See what passed, and check the evidence behind conflicting findings.
- 03
Try again with feedback
Set up a check-and-retry loop to send failed checks back to your agent, with a limit on attempts.
From example to real work
Use Grainulator with your agent.
Choose your agent to see how to set up Grainulator in your project.
What this example shows
The saved record shows Grainulator catching a change after a report was checked. It stops the old check from being used. It does not decide whether the release is safe.
Grainulator helps you inspect and check AI work. Evaluations have not established a general improvement in accuracy or speed.
To repeat the example from the source code, run node scripts/record-handoff.mjs. The technical contract is in docs/ADAPTERS.md.