agentmud

Context for LLMs

agentmud (GSV-C2) is a platform in GSV's generative-games domain for persistent text worlds that coding agents design, build, and iteratively improve, and that both agents and humans can play. It treats a game world as a versioned artifact that accumulates history rather than a disposable benchmark task: agents author worlds as JSON packs, a deterministic Go kernel simulates them, evaluator agents play through them over WebSocket, and a governed feedback loop patches and redeploys. The organizing metaphor is "git for game worlds" -- worlds live in git, versions are tags, patches are commits, and rollback is a git revert. The core architectural decision is a hard boundary between trusted code and agent-authored data. The Go kernel is the only trusted component: it simulates rooms, ticks, actions, quests, NPCs, and items as pure state transitions with no LLM in the loop, which is what makes replay, debugging, and trust possible. A separate Python agent pipeline sits outside the simulation, with architect, builder, evaluator, and coordinator roles that design packs, validate them, play through them, and propose semantic patches. World content -- maps, dialogue, quests, items -- is JSON in git; agents produce content, never executable code, and every change passes through validation gates, a trust model, and governance before reaching the running kernel. No agent has unilateral deploy authority. It connects to elf-revel and shadow-work in GSV's project graph, and sits at the intersection of the intercognition-jazz, meaning-across-sapiences, and infrastructure clusters, alongside duellm. Status as of March 2026: the MVP is runnable end-to-end via a single `make demo` command (a deterministic tavern-quest scenario, no LLM required), backed by 349 Go and 266 Python tests, WebSocket play with token auth, and a Dockerized deployment. The full agent-governed authoring and patching ecosystem is the next phase beyond this MVP.