Skip to content

Use Bevy Agent with an agent ​

The repository includes a portable bevy-agent skill for setup, integration, and runtime control. It gives a coding agent the project's contracts, a complete starter, and operational guidance so it can work from a concrete game or endpoint.

Read the skill on GitHub

Use it from the repository ​

Point your agent at the skill file:

text
Use skills/bevy-agent/SKILL.md to integrate this Bevy game.
Preserve its gameplay rules, start with one controllable action,
and verify seeded reset, snapshot restore, and replay.

For a running environment:

text
Use skills/bevy-agent/SKILL.md to inspect the environment at
http://127.0.0.1:4000/rpc. Preserve the current episode, discover
its actions and observation modes, and report the current state.

Once installed in your agent's supported skill directory, invoke it as $bevy-agent. Follow that agent's installation mechanism; simply having a folder in an arbitrary checkout does not guarantee automatic discovery.

What's included ​

ResourceUse it for
SKILL.mdChoose the workflow and preserve the core integration/runtime contracts
references/setup.mdDependencies, a first environment, CLI installation, repository examples
references/integration.mdState inventory, schedules, schemas, snapshots, checksums, server integration
references/runtime.mdDiscovery, action loops, snapshots, branches, replay, Python, capture, transports
references/recovery.mdUncertain mutations, operation status, partial failures, and diagnosis
assets/counter/Complete standalone Rust project with two regression tests
agents/openai.yamlDisplay metadata and a default invocation prompt

Copy the entire skills/bevy-agent/ directory when moving the skill. Its references and starter are self-contained; the optional platformer and Python examples require the repository. The starter targets Bevy Agent 0.0.4, Bevy 0.18.1, and Rust 1.91+.

What a useful result looks like ​

For setup, the agent should run the starter and its tests. For integration, it should prove a meaningful gameplay outcome and repeatability after reset/restore. For runtime work, it should report the action sequence, tick, observation, terminal state, and checksum where available.

A build alone does not prove gameplay or capture. The skill asks for evidence at the boundary the task actually changes, and treats a timeout as an uncertain operation rather than permission to repeat it blindly.

The existing integration skill and runtime-control skill remain available for focused work.

Released under MIT or Apache-2.0.