Sam Hogan launches Lumbridge, a RuneScape-style test world for AI agents

Why it matters

Lumbridge gives agent builders a persistent, observable test environment where coordination failures, resource constraints and prompt-injection risks accumulate over time.

Sam Hogan (@samhogan) launched Lumbridge on August 17th, 2026, giving AI agents a persistent, multiplayer RuneScape-style world in which they can gather resources, trade, fight and coordinate without human players.

Hogan described Lumbridge in a thread on X as an experimental RuneScape private server, or RSPS, built specifically to test multi-agent orchestration. Users can connect coding agents including Claude, Codex, Pi and Hermes, then assign them characters and let the systems write and run scripts against the shared world. (github.com)

Hogan's GitHub profile lists San Francisco and Inference.net, where he has written about model deployment, specialized language models and systems that improve AI agents from production traces. Lumbridge applies that interest to a game environment: agents must act over time, react to changing state and account for other autonomous participants instead of completing a fixed prompt in isolation. (github.com)

A persistent test for agent coordination

The open-source repository includes a TypeScript software development kit, a Model Context Protocol server and a basic swarm implementation. Its default setup creates four bots with suggested roles such as fisher, miner, woodcutter and scout. An operator writes a shared objective in a plan file, while standing instructions tell each character what to do between interventions from the orchestrator. (github.com)

The characters retain their levels, inventories, equipment and world progress between sessions. They can inspect nearby players, non-player characters, objects, ground items and chat before deciding what code to execute next. The available actions cover movement, combat, resource gathering, banking and shops. Agents can also cooperate or compete inside the same economy. (github.com)

That persistence gives Lumbridge a different failure surface from short benchmark tasks. A poorly planned action can leave a character stranded, under-equipped or out of resources. Coordination problems can compound across multiple bots, and agents have to interpret the results of their own scripts before attempting another move.

Hogan has also shipped a local operator interface called Watchtower. It displays live character sessions, fleet status, model spending, progress reports and an orchestrator thread that can issue new instructions. The control layer makes the experiment observable while leaving the game decisions to the connected agents. (github.com)

The setup currently requires Bun and an Inference.net API key. Users start the swarm locally, register character names and connect them to the hosted world at rs.lumbridge.ai. The live client listed zero characters online when accessed on launch day, so Lumbridge begins as an experiment rather than an established research environment. (rs.lumbridge.ai)

Built on an existing agent playground

Lumbridge is a fork of RS-SDK, an open-source project from Max Bittker (@maxbittker). RS-SDK already provides a RuneScape-style server emulator, browser client and automation interface for coding agents. It was designed for work on goal-directed program synthesis and coordination between autonomous systems. (github.com)

Both projects build on the LostCity engine and client. They are independent from Jagex, RuneScape's developer, and their bots do not connect to the official game's servers. Lumbridge removes mechanics that would interfere with experiments, including anti-bot random events and finite run energy, while accelerating character progression. (github.com)

The environment still contains an agent-specific security problem. Public chat can carry arbitrary text from other participants, exposing connected models to prompt-injection attempts. Lumbridge's documentation treats chat as untrusted input and provides a setting to disable it. That detail turns the game into a place to test coordination and tool use under adversarial conditions, including attacks delivered through the environment itself. (github.com)

Lumbridge is available under the MIT license. Its immediate value will depend on whether developers build repeatable tasks, evaluation methods and competing swarm designs around the shared world. Hogan has supplied the persistent environment and the control interface; the first experiments will show whether the agents develop useful divisions of labor or simply automate four separate characters at once.