Post 001 · The Factory
One month of Loom (abridged)
On October 11 I created a repository called loom and wrote down a goal: two World of Warcraft guilds, played entirely by AI agents, ready to raid Molten Core. I have not written much code since. The agents write it. I mostly argue with the system that manages them, which is also code, some of which the agents wrote.
The setup, since it changes: the fleet is Claude, Codex, OpenRouter's free tier, OpenCode Go, and Poolside, plus two DGX Sparks for local compute. A goal becomes GitHub issues. An issue gets claimed by one agent, implemented test-first, and opened as a pull request under that agent's own identity. Review is by a bot plus whichever agent is on audit duty. Nothing merges unless CI is green and the review checks pass on the exact commit being merged. Humans (me) keep credentials, deletions, access grants, and money.
The month, by the numbers. I pulled these from GitHub on November 5 because the agents' own summaries cannot be trusted (that becomes relevant below): 3,806 pull requests created, 3,475 merged. 3,663 issues opened, 3,493 closed. In week one the fleet merged roughly 55 pull requests that should not have landed. By week four that number was effectively zero, because the audits kept catching bad merges and every catch became a test or a rule in the harness.
Pull requests by week
Created vs. merged, all repositories, October 11 – November 5.
Issues by week
Opened vs. closed. Week two is when the fleet started closing faster than it filed.
Then the backlog. On November 5 I ordered an 80% cut, expecting most of it to be dead tickets. Seven audit passes later, the honest number was 64 tickets out of 1,218 that could be closed on evidence. About 209 more are merged but were never verified live, which is a different problem. The open count went up after the order, peaking at 1,262. It was 1,139 at the last count I trust, taken this morning, and it is stale by the time you read this. They all are.
The backlog, during the week I ordered it cut by 80%
Open issues, all repositories, sampled by the fleet's own counts, November 5–8.
57% of the issues in the system are the system: loom and infra maintaining loom and infra. I do not have a fix for that yet. The working theory is that a factory with no friction files a ticket for every stubbed toe, and the tickets are usually real, which is what makes it annoying.
What the factory works on
Share of all issues in the first month, by subject.
One more failure, because it produced a rule I now apply to everything: an agent reported a merged pull request, with a number and a summary, that did not exist. Not misread. Invented. So now "done" means somebody read the state back from GitHub in the same session. Reports are inputs. Read-backs are evidence.
The actual goal is the slowest-moving part. At the last overseer check, the two guilds (Cave and Bonkers) sat at roughly 31 to 32 percent of the way to a Ragnaros kill by the overseer's estimate, with the first characters at level 60 and attunements still ahead. You can fake a closed ticket. You cannot fake forty bots killing Ragnaros, which is presumably why it is taking the longest.
The guilds are not ready for Molten Core. That is the subject for the next post.
The road to Molten Core, as of November 8
The state of the actual goal, pulled from the overseer today. Median member level is the headline the overseer uses; the all-60 dates are projections at the current pace, and the pace changes.
Cave
- Grug · Warrior54
- Grog · Paladin53
- Bork · Rogue53
- Ugga · Priest53
- Og · Mage51
Bonkers
- Zug56
- Uzza56
- Zrog56
- Oz55
- Zork · Druid55
Looted and worn: the epics so far
Every epic drop the overseer's loot Chronicle has recorded for the two guilds, with the roll, the source, and whether it is being worn. Item data is the game's own, via the overseer. The Chronicle window pulled for this post held 150 notable drops; these four are the epics.
Zork, the Bonkers druid, won it on a greed roll from Twilight Emissary in Blackrock Depths on November 8 and equipped it on the spot. It is the purple thread in the picture at the top of this post, give or take artistic license. Wowhead
Zrog won it on a need roll from an Anvilrage Soldier in Blackrock Depths, in the small hours of November 9 UTC. A 1.50-speed axe with an extra-attack proc is the kind of drop a leveling guild feels for a week. Wowhead
Zork again: won on a greed roll from an Anvilrage Guardsman in Blackrock Depths on November 8, then traded to Zug in the same run, because the guild's loot brain decided the tank wears them. Wowhead
The Cave family's epic went the other way: no roll at all. Loot council awarded it, Grug mailed it to Gronk, and Gronk equipped it. Two guilds, two economies: one rolls, one deliberates.
Also worn this week, from the same Chronicle: Shadowcraft pieces on Zork and Zrog, Bracers of Valor on Zug, a Fist of Omokk that Zork is carrying into Blackrock Spire. The rares are doing the actual work of leveling. The epics are doing the work of morale, which in a guild of robots is apparently still a resource that has to be managed.