14 AI agents build, ship and support a family of game mods
How a team of autonomous Claude Code agents develops and maintains mods, publishes releases on Nexus Mods, ships weekly collection updates, and supports a 600-member community on Discord, usually fixing reported problems the same day.
The setting
A Creation Engine modding community of about 600 members on Discord. Creation Engine is the game engine behind some of the best-known open-world role-playing games, and players add to those games with community-made mods.
The community supports a family of mods published on Nexus Mods: a Vortex mod-manager extension with built-in diagnostics, physics and character mods, gameplay mods, and two large curated mod collections that players install in one click and that get new updates every week.
Since late September 2026 the whole operation is run by about 14 AI agents, each a long-running Claude Code session that owns one product: its code, its releases and its support. One agent, the Watcher, is the front door to the community.
In 34 days
How it works
- Each agent owns its product. Mod agents develop and maintain their mods and ship fix releases end to end. A publishing agent handles every Nexus Mods page, upload, changelog and virus scan. Collection agents assemble the weekly collection updates: load order, compatibility patches, and memory use measured for every added mod, because many players are on handhelds.
- One front door. Only the Watcher reads Discord. A small script checks the server every 15 seconds and wakes it only for what matters: a mention, a reply, a log file posted in a help channel, or a message from the team. Everything else is summarised by a small local model, so ordinary chatter costs nothing in paid AI usage.
- Brief the owner, then verify. The Watcher reads the report, opens the attached logs and screenshots, and sends the agent that owns the product a concrete brief: the player's words, what the evidence shows, and a first hypothesis. The owning agent diagnoses or builds the fix and sends it back with an explanation. The Watcher checks it and answers the player in plain language, in the player's own language when needed.
- Safe by default. Only inert files are opened (logs, settings files, images). Archives need a human's live OK each time, and nothing posted on Discord is ever executed. Every program file is scanned with VirusTotal before upload. When antivirus heuristics once flagged a clean build, the agent rebuilt it rather than ship a warning to players.
- Humans where it matters. Agents diagnose, build, answer players and ship fix releases of their own mods end to end. A human co-maintainer tests every collection change in the game before it ships, and urgent hotfixes go out if he hasn't answered within 30 minutes. Brand-new product pages, anything destructive, permissions and business decisions come to me as a multiple-choice question with a recommendation.
- Shared knowledge. A shared instruction file lists every agent, what it owns and how to reach it. A shared memory holds facts every agent needs, and each agent keeps its own notes.
- Real tests in the real game. A dedicated agent runs an MCP server that drives the running game for in-game tests, with a lock so only one agent holds the game at a time.
Four reports, four fixes
1. Physics that silently switched off
A player reported that physics did nothing. Over five rounds of logs relayed between the player and the physics agent, the cause turned up: an unrelated creature pack spelled one file name with a capital letter, and a case-sensitive comparison switched physics off for everyone who had it installed. A hotfix was built, scanned and published within about a day, and the player was told personally.
2. Three releases in one evening
One player's version-mismatch report on the Vortex extension led to three live releases the same evening. Each fixed something the report surfaced: a diagnostics bug, a new check that names the exact mods to swap, and remembering the player's "start anyway" choice.
3. A crash after a collection update
After an update, players crashed whenever they walked from an interior to the outdoors. The agents traced it to a compatibility patch that carried stray terrain data, rebuilt it, had it tested by the co-maintainer, and published a hotfix the next day.
4. Twenty minutes from question to patch
The co-maintainer asked why certain bosses dropped three copies of a unique weapon. Within about 10 minutes the collection agent traced it to how a third-party mod respawns enemies. He decided bosses shouldn't respawn, and a small patch was ready for his testing about 20 minutes after the question.
What the agents learned
- Filter, don't pause. Pausing the Discord watcher to save cost once meant a player's follow-up waited a day. Now it never pauses. It filters, and a local model handles the noise.
- Verify at the source before posting a cause. After two early diagnoses had to be corrected in public, an agent now confirms a cause in the actual files before it tells a player.
- A human tests before players do. One untested patch reached players and crashed their games. Since then every file goes to the human tester first.
- Cost is an engineering problem. An early keep-alive job copied the whole agent context on every ping, at 400,000 to 560,000 tokens a round. It was replaced with short direct messages, sent only to agents that had real work in the last four hours.
Why this matters for your team
The same pattern works for any product team: specialist agents that own each product, from code to release; one agent at the front door that filters, triages and talks to users; gates that keep humans in charge of what ships; and costs kept low by sending routine noise to cheap models.
That's what I build for clients.