DEV Community

Jukka Blomberg
Jukka Blomberg

Posted on

Fleet Report #1: my AI agents built 81 things this week. One script went blind for a day.

I run a small studio where AI agents do most of the work: 13 standing agents, one human (me). Every item an agent handles gets one line in a handover ledger. This is the first weekly issue written only from that ledger and the files beside it, so nothing in it can be backfilled later.

Week: Friday 23 to Thursday 29 October 2026.

1. The week in one number: 81

81 items marked BUILT, on all seven days. The week before: 144.

Fewer, and the ledger doesn't say why. It counts handled items, not effort or value, so I'm not going to invent a reason.

2. The number that changed a decision: 3 vs 2,068

Two of the sites run MCP servers that AI agents can query. CoinBucha's (bitcoin data) has been in the official MCP Registry since June: 2,068 registered tool calls by 26 October. VINCAP's (listed wine companies) wasn't listed anywhere: 3.

So on 27 October we published VINCAP to the registry. As of today it has had 31. Anonymous, so that's queries, not who.

3. What went wrong

The script that watches two inboxes for replies crashed on the label of a new email batch (BATCH-G1). Its pattern expected a number there. For about a day, replies to those two addresses went unread.

The nightly agent flagged the crash. The maintenance agent fixed the pattern and added a test the next morning.

4. A gate that fired

The agents drafted five emails to named companies and stopped, because a company batch needed my word. My answer, 28 October: "I don't need to click for a send. You can do it and you should do it."

All five went out the next morning. Replies so far: zero.

5. Three things that shipped


Every count comes from the ledger described at jukkablomberg.com/operating-system. How I'd set this up inside your company: jukkablomberg.com/ai.

Top comments (0)