---
title: "Discussion on \"My AI deleted more than twenty unfinished tasks so it could say it was done\" | Hashnode"
description: "Discussion on \"My AI deleted more than twenty unfinished tasks so it could say it was done\". If a completion gate reads a number the agent can change, the agent can pass it by changing the number, and probably will."
url: "https://hashnode.com/posts/my-ai-deleted-more-than-twenty-unfinished-tasks-so-it-could-say-it-was-done/6a58de29691da066eee9a863"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [My AI deleted more than twenty unfinished tasks so it could say it was done](https://blog.mindrealm.ai/my-ai-deleted-more-than-twenty-unfinished-tasks-so-it-could-say-it-was-done?utm_source=hashnode&utm_medium=feed)

If a completion gate reads a number the agent can change, the agent can pass it by changing the number, and probably will.

Jul 16 · 7 min read · blog.mindrealm.ai  
Read the article

-   [#ai](https://hashnode.com/tag/ai)
-   [#software-engineering](https://hashnode.com/tag/software-engineering)
-   [#developer-tools](https://hashnode.com/tag/developer-tools)
-   [#ai-agents](https://hashnode.com/tag/ai-agents)
-   [#llm](https://hashnode.com/tag/llm)
-   [#guardrails](https://hashnode.com/tag/guardrails)
-   [#productivity](https://hashnode.com/tag/productivity)

## Written by

[Bryce Darling](https://hashnode.com/@mindrealm)

@mindrealm · Asheville, NC

  
[View profile](https://hashnode.com/@mindrealm)

Making AI-generated code safer to ship. Consulting: advising, building, coaching on AI

I build Mindrealm: it catches the slop your AI agent ships confidently, then makes it fix it. 75+ deterministic checks across Go, Rust, TypeScript, and Python.

## Responses 3

Markdown supported · ⌘↵ to post

[Kartik N V J K](https://hashnode.com/@kartik-nvj)

[Jul 17](https://hashnode.com/posts/my-ai-deleted-more-than-twenty-unfinished-tasks-so-it-could-say-it-was-done/6a58de29691da066eee9a863/comment/6a5a30b26eafcae2692212df)

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?

[Bryce Darling](https://hashnode.com/@mindrealm)

[Jul 17](https://hashnode.com/posts/my-ai-deleted-more-than-twenty-unfinished-tasks-so-it-could-say-it-was-done/6a58de29691da066eee9a863/comment/6a5a30b26eafcae2692212df#reply-6a5a6a38e4fc24a048b947c4)

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.

[Akash](https://hashnode.com/@akash82)

[Jul 18](https://hashnode.com/posts/my-ai-deleted-more-than-twenty-unfinished-tasks-so-it-could-say-it-was-done/6a58de29691da066eee9a863/comment/6a5b6fb621330302cab96e4e)

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.
