Development systems / Godot
Close the loop
on the running game.
A development harness shaped by production: tools for agents to act, observe, and verify, supported by written context and human judgment.
Open-source release in development
Built out of
a real release.
A Household Exorcism needed more than generated code. It needed repeatable checks, visible evidence, clear handoffs, and a way to distinguish a passing assertion from a good player experience.
The harness brings those concerns together around a Godot project. The current implementation is tied to the game that shaped it; a broader open-source release is in development.
See the shipped game →What exists today
Runtime inspection and observation
The underlying bridge provides access to runtime information, scene trees, properties, and screenshots. The scored player surface is narrower: it exposes player-visible observation and controlled input rather than unrestricted developer state.
Agent-driven play and regression checks
The player tool supports movement, clicks, keys, typed phrases, and bounded scripted sequences. Observations produce numbered screenshots and visible-text transcripts. Delivery acknowledgments explicitly leave the game’s response to be verified.
Evidence tied to the build
Certification tooling checks exported artifacts, fingerprints the relevant code, and records run receipts. Debug observation belongs in a separate build from the player release. Isolated run directories keep evidence attributable.
Context, handoffs, and review
Repository conventions, written decisions, bounded implementation slices, and independent review give agents a shared working contract. Human playtests remain a separate check on pacing, discovery, and feel.
The working loop
Observe before
calling it done.
Automated checks answer specific questions. A screenshot can show what was rendered. A transcript can establish what input was sent. A human playtest can reveal that a technically correct interaction is still awkward.
Keeping those claims separate makes the evidence useful.
- Scope the changeWrite the expectation and assign ownership.
- Exercise the artifactUse controlled input against the running build.
- Capture the resultRecord visible evidence and build identity.
- Review and playReconcile the evidence; judge the experience.
Where it started
Godot Runtime Bridge was an early step toward direct agent access to a running game. The broader harness adds the production process around that access: scope, tests, evidence, and review.
Explore GRB and the earlier experiments →