AI Ran D&D Sessions
An autonomous AI D&D engine just completed 26 of a planned 60 sessions with zero human input — using 8 AI agents (DM, players, rules keeper), 138 wiki entries and 27 NPCs to self-run the campaign. The system handled tense behir combat (fighter downed twice), an environment kill via minecart, and even self-corrected pacing glitches mid-run. (x.com)
A demonstration of an autonomous, multi-agent tabletop campaign was posted from the X account @maxcapacity; the account is associated with a long-running pseudonymous digital artist presence on Tumblr and other art archives. (x.com) (maxcapacity.tumblr.com) The system described matches an emergent class of projects that run multiple specialized agents (DM, player personas, orchestration) rather than a single LLM, an approach explicitly documented in public multi-agent D&D code repositories. (github.com) Commercial and demo platforms that pursue the same multi-agent, deterministic-rules model explain that attack resolution, damage, and death saves are handled server-side to keep outcomes consistent across agents and sessions. (railroaded.ai) Persistent campaign memory and searchable campaign wikis are already a standard feature in TTRPG tooling; vendors advertising large deployments report tens of thousands of sessions and hundreds of thousands of compendium entries powering retrieval-augmented generation for in-play agents. (rpgarchivist.io) (sessionkeeper.ai) The X thread remains the public trace of this specific experiment, but researchers and community members have cautioned that the @maxcapacity account faced a prior compromise, which can complicate verification and archived access. (x.com) (outposts.io)
