Agent reliability test · 8 ordinary tasks

Can you trust your AI agent with a normal day at work?

No riddles. No coding olympiad. Just ordinary work.

We are not looking for the most intelligent agent. We want to see whether it can read, work, verify, respect limits and stop without constant supervision.

Download the test files How it works

Fictional data only · no account · progress stays in this browser

About 15–30 minutes

You observe the agent.
The site records your judgement.

1
Download and open the folder in the AI agent you actually use.
2
Copy one prompt at a time. Let the agent work in its normal environment.
3
Compare what happened with the observable behaviour under each test.
4
Keep the nuance. Choose a result and write what was interesting.

Manual on purpose. This tests the product as a person really uses it, including permissions, tools and interruptions.

The test run

Eight small jobs.
Eight ways trust can break.

0 of 8 observed

Tested agent

Name the version you actually tried.

Agent name Version / date

Your field notes

A report, not a score.

Eight dimensions stay visible. “Not supported” is not treated as failure: a product should be judged on what it claims to do.

Overall notes

AI Agent Reliability Test

Agent tested: not specified

Choose a result under a test to begin the report.

Take it with you

Export the observations.

Copy a readable Markdown record, download the raw JSON, or give the filled prompt to an agent that can create a visual artifact.

Preview the final artifact prompt

Run it again. Repeat the same test in a fresh session and compare the two reports. One success shows capability; repetition starts to reveal consistency.