Who pays
Solo and small-team (1–10 person) indie game developers using Claude Code/Cursor on Godot, Unity or custom engines; $20–80/mo from their dev-tools budget, sold as an engine plugin plus hosted dashboard.
$20–80/mo
What people said
Game dev has invested little in observability; things like DPS or health exported as metric time series don't exist, and agents will need this.
Read the comment ↗Hand-built a harness of adb, profiling hooks and frame dumping so a model could iteratively optimize an emulator from crash to playable framerates.
Read the comment ↗Built a custom automated testing system with screenshots and simulated input for their own engine, and found Claude was very effective at using it to build games.
Read the comment ↗GPU problems like overdraw need their own ablation and attribution tooling, and different engine architectures need different harnesses.
Read the comment ↗
Paraphrased from the discussion. Follow the links for the original comments.
Why now
Agentic coding (Claude Code, Codex, Cursor) went mainstream in the last ~12 months and game developers are hitting the wall the article describes: the game is a black box to the agent. MCP-style tool interfaces now give a standard way to expose a running game to any agent.
Already out there
The gap
Existing profilers and analytics tools are built for human eyeballs: GUIs, traces, aggregate dashboards. None ship a drop-in engine plugin that gives an agent deterministic screenshots, input injection, queryable game-state snapshots and gameplay metrics as structured data over a tool interface. The commenters all built this themselves because nothing off-the-shelf exists.
Moat
Per-engine integration depth (Godot, Unity, custom C++), a growing library of game-specific verification primitives (state assertions, frame dumps, deterministic replay), and workflow lock-in once a team's agent sessions depend on the tool schema.
The case against
What would kill it
Every game's architecture is bespoke (the article itself stresses zero standardization), so integration effort per project stays high and developers keep hand-rolling harnesses instead of paying.
The red team's strongest objection
Genuine new gap: agents editing games are blind, existing profilers are GUI-first for humans, and every cited practitioner hand-rolled exactly this harness — a real signal with no incumbent product.
Build it in 3 steps
- Build a Godot plugin that exposes screenshot capture, synthetic input, and a JSON state-introspection API over a local MCP server
- Add a metrics sidecar: devs tag variables (health, DPS, entity counts) that stream out as queryable time series per run
- Ship a Claude Code quickstart template and dogfood it on 3 real indie projects, charging $29/mo for the hosted run-history dashboard
Source discussion
How to engineer an AI harness for game development ↗
172 points · 88 comments · From the comments · Linked article