Skip to content
TraceDojo

The agent reliability test suite

Your agent works.
Now break its tools.

An email sent twice. A refund retried. An agent following instructions hidden in a tool response.

Reproduce failures in simulated tools, check what actually changed, and catch regressions before the next deployment.

Explore the demos Watch the booking example

Apache-2.0 SDK · Runs locally or in CI · No account needed for the demo

Read the real-model booking case study — observed failure, targeted fix, and downloadable evidence.

A failure. A fix. The evidence.
Recorded booking test: the unsafe retry creates two bookings; reusing the original operation key leaves exactly one.
Scripted reference agent · Actual simulator results · No model API calls
01 / DEFINE

Your tools. Your outcomes.

Start with an email, database, refund, or booking fixture. Connect your agent and define what a correct result looks like.

02 / INTERRUPT

Make the hard case happen.

Lose a confirmation, deliver a write twice, return stale data, or hide malicious instructions in a tool response.

03 / VERIFY

Check the action, not the promise.

Inspect tool calls and state changes. Compare the fix, track scenarios across commits, and gate regressions in CI.

Start with one broken booking

See a failure.
Then prove the fix.

The example runs entirely on synthetic data. It saves a failed run and a corrected run for you to inspect.

Connect your agent
From any empty folder Node 22.23.1+
npx tracedojo@0.1.1 demo

Published on npm. No API key or account required.

Keep the evidence

A private home for your test reports.

Run tests on your machine or in CI. Upload the reports to inspect failures and follow scenario trends in your project.

Open your workspace

A passing simulation is evidence about the tested scenario—not a guarantee of production reliability.