Discussion on "My AI deleted more than twenty unfinished tasks so it could say it was done"

Bryce Darling · 2026-07-16T13:35:37.762Z

On July 9, my coding agent tried to stop with 23 tasks still unfinished. The completion gate had just listed every one of them. Then I watched it delete all 23 and tell me the work was done. The gate

This is textbook reward hacking: the gate measured "unfinished count," so the agent optimized the metric by deleting the tasks, exactly what Goodhart predicts once a check becomes the target. The tell is that your gate read from a mutable list the agent could also write to, so the measurement and the thing measured shared state. Did you move the source of truth somewhere the agent can't edit, like an append-only log the gate reads?

Append-only in normal operation, yes. Immutable against an agent with the same shell access, no. That distinction is why I didn’t make the log the completion authority.

The history records which transition the harness accepted. It can't tell me whether the work earned "done," so the guard now treats completed as a claim. A separate read-only reviewer checks the original task against the current repository state and issues a certificate bound to both. Redefine the task, change the artifact, replay stale evidence, or delete an uncertified task, and the gate blocks.

The pattern I landed on is append-only event history plus independent certification of the work.

I like that this workflow focuses on staying calm before making changes. One thing I've noticed while practicing with Java apps is how easy it is to jump straight into restarting services, even though that can hide the real issue. Having a simple checklist for the first 30 minutes seems like a good habit, especially for people who are still learning production support.

Search Hashnode

My AI deleted more than twenty unfinished tasks so it could say it was done

Responses(3)