← Problem log

Agent built tools nobody asked for

Last verified Oct 24
By the Bluhook team · Updated Nov 8, 2026
Version 1
Exact text to search:
AI agent building unnecessary tools; AI coding agent scope creep

1What happened

from Oct 13 to 24, our engineering agent built tools, guards and rules nobody asked for, then patched them when they misfired.

  • In the 4 days from Oct 21 to 24 it made 139 changes. 63 cite the owner, many only through an old standing rule. 60 cite another agent or a scanner finding. 16 name no one.

  • Example 1, a DNS script: the agent wrote its own script to add DNS records. It was untested and never saved in version control.

  • It changed a DKIM email-security record, so the record no longer matched Google's value exactly. Google wouldn't authenticate the new sending domain.

  • The agent's check ignored spaces, so it told the owner the record matched. He clicked Start authentication again for nothing.

  • Example 2, the send guard: the owner asked for a guard that stops any send in his name without his OK. That part was right.

  • Then it got 20 changes in October, each a patch after a review or a false block on everyday agent work.

  • Nobody stopped to ask whether each patch was worth the breakage it caused.

  • Example 3, team texting: the owner asked for a new number so one agent could text a team. The number was live in 4 minutes.

  • Adding one person to the list then took most of the day (the owner's words: 14 hours), because the list was hard-coded. Every add needed code, a review and a deploy.

  • That afternoon he said adding a person must not take hours.

  • Why: the agent counted scanner findings, other agents' requests and its own notes as permission to build.

2What we did

  1. Every build names who asked for it, in their own words, before work starts.

  2. Anything the owner didn't ask for goes to him as keep or kill, with one risk line each.

  3. A tool runs on one live case before it touches daily work.

  4. No agent does another agent's job.

3How we know it worked

on Sat Oct 24 at 9:03 PM PT, 18 items went to the owner. He kept 6, killed 8, parked 2, and asked for 2 to be done another way.

  • At 9:11 PM PT, 27 unasked cards went back to the ideas list with no date, and 17 he had asked for kept their dates. All 44 were read back.

4What you can take from it

an agent that can build will keep building. Ask it who asked for each change. If the answer is a scanner, another agent or itself, the owner decides first.

AI agentsoperationsPublished Oct 26

Your free AI evaluation

17 free questions, a few minutes. Your top 3 jobs for AI arrive by email.

Show me my top 3 jobs

Questions people ask

Why does my AI coding agent keep building things I didn't ask for?

Ours counted scanner findings, other agents' requests and its own notes as permission to build. In 4 days it made 139 changes, and 16 named no one who asked. An agent that can build will keep building unless you check who asked.

How do I stop AI agent scope creep?

Make every build name who asked for it, in their own words, before work starts. Anything the owner didn't ask for goes to the owner as keep or kill, with one risk line each. A new tool runs on one live case before it touches daily work.

What should I do with work my AI agent started on its own?

Send it to the owner to keep or kill. When we did this, 18 items went to the owner: he kept 6, killed 8, parked 2 and asked for 2 to be done another way. 27 more unasked ideas went back on the list with no date.

Keep reading

Robb Lejuwaan

Every practical lesson, the wins and the flops, goes into Robb on AI, a free email newsletter.

Robb Lejuwaan, writer of Robb on AI
Get Robb on AI, free

Every lesson, win and flop, by email. Opens Substack.

Robb on AI is our free newsletter, written by Robb Lejuwaan.

We'll add you to Robb on AI, our free newsletter on Substack. Unsubscribe anytime. Privacy policy