Pimp My IDE / garage dispatchBack to the garage
13 Oct 2026 / game development / agent feedback

Give the agent a state it can hit twice.

Game code is only half the machine. A useful agent loop also needs a fixed scene, a typed action route, a narrow visual witness, and a profile it can read without guessing at a chart.

Fixed fixtureCommand replayScoped witnesses

The game must expose more than source code.

Kyle Kukshtel argues that coding agents struggle with games because much of a game's state lives outside plain source files. His proposed route starts with fixed randomness, saveable state, serialized commands, scoped screenshots, image diffs, and text-readable profiling.[1]

That is one developer's architecture argument, not a universal game-engine rule. The durable point is smaller. If a change depends on scene state, timing, assets, or runtime performance, the review loop needs a stable way to reach that state and a witness matched to the claim.

Do not ask a screenshot to prove simulation logic, or ask a unit test to prove the frame looks right.

Split control from observation.

A serialized command log can drive the same fixture again. It does not prove the replay reached the same state. Save the build identity, fixture digest, seed, command list, clock policy, and final state digest. If the game includes nondeterministic physics, networking, or platform behavior, record that limit instead of calling the replay exact.

The witness should follow the claim. Use a cropped frame diff for a visual target. Use state values or invariants for simulation behavior. Use a text profile for CPU or allocation claims. Keep the full capture available for a human who needs to inspect the path behind a number.

Tool access is now reaching these witnesses.

JetBrains says Rider 2026.2 can give agents coverage data, profiler snapshots, refactoring support, and game-aware Unreal workflows. Its dottrace-analyze skill reads a supplied snapshot and traces hot paths into code.[2] These are vendor claims about Rider's current release. They do not establish that every game project, engine, or profiler export will work well with an agent.

Chrome's October DevTools recap shows the same movement on the web side. The DevTools MCP server added direct heap-snapshot queries, optional console stack traces, source inspection, and controls that can disable JavaScript evaluation. Chrome also added a CPU performance-tier override over the DevTools Protocol for reproducible performance checks across calibrated hardware tiers.[3]

A narrow port beats a giant dump.

More telemetry can make the loop worse if the agent receives an entire scene tree, full-resolution frame, profiler database, and console log on every turn. Give each task the smallest useful fixture and witness. Preserve the raw artifact outside the prompt so a reviewer can inspect it later.

Start with one expensive bug. Save the state immediately before it. Route one typed command through the simulation. Capture one relevant frame region and one text metric. Run the same packet before and after the patch. That is enough to test whether the port helps before redesigning the game around it.

Interactive makeover / game test packet

Wire the witness bay.

A normal checklist says whether tooling exists. This bay shows whether one replay packet has an unbroken route through state, action, frame, and profile records. Selecting a section only adds it to the packet structure.

Load stateReplay commandCompare frameRead profile

Packet sections

Use square checkboxes because these records are independent. A later section cannot repair an earlier gap.

Include in the request

Test packet

No packet section is selected. No replay is specified.

State
Command
Frame
Profile
4 packet sections open.

Select the records this test request must collect.

Sources read / 13 Oct 2026

Source log

Open the evidence boundary
  1. Kyle Kukshtel, "Your Game Is A Harness", published October 10, 2026 and read October 13, 2026. We used the author's first-person game-development argument for fixed randomness, saveable state, commands, image diffs, standalone simulation, and text-readable profiling.
  2. JetBrains, "Rider 2026.2: IDE Intelligence for AI Agents, Faster Performance, and Spectacular Game Dev Updates", published July 22, 2026 and read October 13, 2026. We used it for the vendor's description of agent access to profiler, coverage, refactoring, and game-aware workflows.
  3. Chrome for Developers, "New in DevTools - October 2026", published September 22, 2026 and read October 13, 2026. We used it for current DevTools agent controls, heap queries, console traces, and CPU performance-tier overrides.
  4. Hacker News discussion for Kukshtel's post, item 50072120, confirmed through the Hacker News API on October 13, 2026. It was a discovery source. Technical claims above come from the three pages listed first.

Evidence boundary: we read the full game-development essay and the two vendor release posts. We did not inspect Kukshtel's private game code, run Rider's agent skills, connect the Chrome DevTools MCP server, or reproduce a game fixture. The witness bay drafts a test packet. It does not load a game, execute commands, compare frames, profile a process, or approve a patch.