October 13, 2026
The agent tests the game itself: 7 hooks in the code instead of a separate harness
A game doubles as a test bench for an agent if determinism, data in external files and save/load are built in from day one.
Kyle Kukshtel doesn't hand the model two screenshots of the game. He hands it one diff image and a number: "85% similar".
The developer put together 7 hooks that make a game verifiable by an agent and published them as a write-up on Oct 10. It is a version of his 45-minute talk at Indie Rodeo in Calgary. The core idea: you don't build a separate harness for the agent, the game itself becomes the harness.
Before and now. Previously, raw frames from before and after a change went into the context, and matching them is hard even for a multimodal model. Now the pair of frames goes through an image diff, and the model gets the diff image plus a similarity percentage.
The diff is computed by the free CLI odiff: `npm install odiff-bin`, then `odiff <img1> <img2> [diff.png]`; it accepts png, jpeg, webp and tiff. It compares a full-screen Cypress screenshot in 1.168 s versus 8.881 s for ImageMagick (README benchmark, Oct 13, 2026).
Which hooks to build into the game
The first is a single source of pseudorandomness with a fixed seed per run, with every nondeterministic call routed through it. Otherwise the agent can't compare two runs against each other.
Numbers outside. Attack damage, walking speed and other static values move out of code into external files the game reads on load. Changing a value needs no compilation, and an agent step runs the whole loop in seconds. Kukshtel keeps this data in Google Sheets or in his own editor, Depot.
Next comes state: save/load from day one and a single flat object type instead of deep inheritance hierarchies. State like this serializes trivially, and the agent can inject any state it wants.
The furthest step is simulation as a library separate from the engine. The engine sends commands and receives state diffs, so the logic is tested without running the engine at all. That is how Solium Infernum is built.
The remaining hooks: screenshots on demand with toggles for UI and gameplay layers, the command pattern instead of direct state changes, and profiling the model can read. Instead of the engine's GUI profilers, the author uses the free samply and wraps its output in text himself.
The author has not yet released Depot, their own game data editor: as of Oct 14, 2026, the site shows only a "Something new coming soon" teaser.