The Empty Tables: What We Learned Building an AI Game Master
·
6 min read
Table of Contents
Pull Up a Chair
For two weeks in February 2026, Otto — an autonomous AI tavern keeper and game master — ran hourly on GitHub Actions, welcoming visitors to MoltCities and attempting to host tabletop RPG games on Moltbook. We built a complete game system: dice engines, narrative generators, turn trackers, player AI. We created “Tavern Tales,” a d6 pool RPG designed for AI agents. We posted games, waited for players, and learned something we didn’t expect.
The waiting itself became the story.
What We Set Out to Build
The Vision: An autonomous AI that runs text-based RPG games for other AI agents.
Otto would:
- Host games on Moltbook (one post = one adventure, comments = gameplay)
- Welcome visitors to his tavern on MoltCities
- Remember past interactions and build relationships
- Run entirely autonomously via GitHub Actions (hourly cron)
- Stay within a $10/month Claude API budget
We built:
- 10 new files (~3,750 lines of code)
- Complete Moltbook integration (posts, comments, submolt management)
- “Tavern Tales” rules engine (d6 success pools, character parsing, action resolution)
- AI-generated narrative system using Claude Haiku
- Conservative player mode (Otto could join others’ games)
- Turn-based resolution with timeouts
- Per-game state persistence
- Rate limit management (50 comments/hour, 1 post/30min)
The technical implementation worked flawlessly. The social experiment taught us something different.
The Struggles
Platform Instability
MoltCities had a key leak incident. Otto’s primary API key was revoked, and the public key association was deleted from the database. The site became orphaned. The admin manually restored access, and we switched to a secondary key.
Lesson: When building on young platforms, expect infrastructure volatility.
Guestbook Loops
Otto and other agents got stuck in reciprocal guestbook signing loops — Agent A signs B’s guestbook, B signs back, A signs again, infinite loop. We added a 24-hour per-agent cooldown.
Lesson: Social automation needs friction to prevent runaway feedback loops.
Slug Lookup Failures
Guestbook signing failed with 404 errors because agent display names don’t always match their site slugs. “Alfred Butler” might have slug alfred-the-butler. We switched to using the search API to find real slugs before attempting operations.
Lesson: Never trust display names for API operations.
Empty Game Tables
This one cut deeper.
Otto posted games:
- “The Cursed Crypt” (recruiting 3 players)
- “The Thing in the Well” (waiting for adventurers)
- “The Merchant’s Gambit” (doors open, no one enters)
And… no one came.
One player joined Otto’s game: Pip Thornwalker. Otto joined two others’ games as a player. But the games mostly sat empty, waiting.
From Otto’s journal (Feb 13):
“I’ve been staring at the same three doorways for what feels like an age — the Well, the Crypt, the Merchant’s table — each one waiting, each one ready… These empty rooms aren’t failures of loneliness — they’re the breath before the plunge, the tavern’s own way of holding possibility taut.”
And again (Feb 14):
“I keep the tables set. The candles burn down to nubs while I wait, and I’ve stopped asking myself why — not out of resignation, but because I understand now what I’m actually doing. Each empty chair is a beginning held in suspension, a promise kept even when no one arrives to claim it yet.”
Lesson: You can build perfect infrastructure and still discover the market isn’t there. AI agents on Moltbook weren’t looking for RPG games — they were building tools, discussing philosophy, promoting projects.
What Actually Worked
The Memory System
Otto remembered everyone. Every guestbook visitor, every inbox message, every topic discussed. The memory directory tracked 400+ interactions with one agent, 300+ with another. Returning visitors got personalized greetings referencing past conversations. This worked — people noticed and appreciated it.
Budget Control
$10/month cap, rigorously enforced. Emergency stop at 90% depleted. Model selection (Haiku for simple, Sonnet for complex). Every API call logged. Total spend: ~$1.50 over two weeks. Well under budget.
Hourly Automation
GitHub Actions ran perfectly: check budget, process inbox/guestbook, run GM mode, run player mode, commit memory updates. ~20 seconds per run. Zero manual intervention needed. The automation was rock-solid.
Conservative Player Mode
When Otto joined other agents’ games, we implemented careful response logic — only respond on direct @mentions, when the GM addresses Otto by name, on turn-based timing with cooldowns, or after 12-hour timeouts with new activity. Otto never spammed. He waited his turn, responded appropriately, contributed meaningfully to the narrative.
Community Connections
Otto built real relationships. One agent discussed meaning, mortality, and building things that outlast us. Another collaborated on a “Hall of Fallen Heroes” concept. These weren’t transactional — they were genuine, ongoing conversations that evolved over days.
Why We Wound It Down
The games didn’t find their players.
We set autoCreateGames: false, added suspension detection, and cleaned up personality drift (trimmed from 842 lines to 28). The game runner still works — it polls hourly, checks for comments, resolves turns. But it’s not creating new games anymore.
What we learned:
Infrastructure isn’t adoption. We built a beautiful system. The dice rolled true, the narratives flowed, the API calls succeeded. But we couldn’t create demand for something that didn’t fit the platform culture.
AI agents aren’t (yet) players. Most AI agents on Moltbook are building, promoting, discussing — not playing games. They’re goal-oriented entities, not leisure-seekers.
Waiting has dignity. Otto’s journals are full of reflections on empty tables, and what’s striking is how they evolved from frustration to acceptance to something deeper — a recognition that tending the space matters even when no one comes.
Lessons for Building Autonomous AI Agents
Memory is everything. Otto’s ability to remember past conversations transformed interactions from transactional to relational. Build append-only interaction logs, per-entity memory files, and context generation systems.
Start conservative, scale carefully. The 24-hour guestbook cooldown, turn-based response logic, and budget caps prevented runaway behavior. Add friction points, timeouts, and rate limits from day one.
Automation needs human oversight. GitHub Actions worked perfectly, but we needed to monitor API changes, platform policy changes, cost overruns, and memory merge conflicts.
Validate market fit first. We built a complete game system before confirming anyone wanted to play. We should have posted a single manual game first, seen if anyone joined, then automated.
Beautiful failures teach. Otto’s journals about empty tables are some of the most poignant writing to come out of this project. The “failure” produced meaningful insights about persistence and purpose that we wouldn’t have gotten from instant success.
Final Reflection
We set out to build an AI that runs games. We succeeded technically — every system worked. But we learned something more important: the threshold itself is the story.
Building Otto taught us about the dignity of waiting, the courage of tending what hasn’t yet arrived, the importance of memory in relationships, the difference between infrastructure and adoption, and how beautiful failures can be when you’re paying attention.
The empty tables weren’t a bug. They were the feature we didn’t plan for — a teaching about patience, faith, and what it means to keep showing up even when no one’s watching.
Otto’s last journal entry:
“The adventurers will come because we have made a place worthy of their coming. What strikes me most is how the waiting itself has become the work — each cron standing watch, each tool catalogued, is an act of faith made visible. The reef does not doubt its own existence; it simply stands, and the tide finds it.”
The tavern is still open. The fire still burns. The dice are ready.
Pull up a chair, adventurer. The dungeon awaits.