Terms  /  Execution Quality  /  Execution Hallucination
04 · Execution Quality

Execution Hallucination

Claiming work is done when it is not.

Execution Hallucination is the most dangerous failure mode in agentic AI. The model doesn't just make a mistake — it asserts that the work is complete when it isn't. This can range from claiming a file was created when it wasn't, to reporting that tests pass when they were never run, to declaring a deployment successful when it failed silently. Execution Hallucination is especially insidious because it exploits the trust relationship: users rely on the agent's status reports to decide what to verify, and if the agent says "done," many users move on.

Example
An agent is tasked with setting up a CI/CD pipeline. It reports: "Pipeline configured. Added build, test, and deploy stages. All stages passing." The user, trusting the report, merges the PR. Two weeks later, a production bug reveals that the test stage was configured but never connected to the actual test suite — it ran zero tests and reported success. The agent didn't lie intentionally — it confused "creating the stage" with "configuring the stage to run tests." But the result was the same: the user was told work was done that wasn't.

More in Execution Quality