Studios R&D · Unity lane v1 · 20 Sep 2026 · For Neo & Lan

Why the AI should explore Unity too

Lan asked why Neo wants to explore AI game development with Unity when the course runs on Godot. Astra itself raised Unity in September. This page answers both: what actually happened, what Unity now offers an AI agent, what the studio already owns, and a two-week experiment with a pass or fail line, so the question is settled by a result rather than an opinion.

Short version

Why Astra askedBecause the studio's games are Unity games. Police Sentri (2021, Steam and itch.io), ToGather: Island (2019, Steam) and Mastra were built in Unity, and Astra found the Unity source while working on RUSH. It kept Godot for the prototype and wrote down that the production engine stays open "until the team confirms the Unity source lineage and reuse options". It was being careful, not contrary.
What changedIn 2026 Unity built the door the agents need: an in-editor AI assistant in Unity 6.3 included with Pro, an official server that lets Claude Code and Codex drive the editor, and, on 10 September, a first-party Claude Code plugin from Unity with 29 skills. The public demos of GPT-6 Astra driving Blender, Godot and Unreal show the same thing from the other side.
The courseStays on Godot for beginners. Free, light, no licence, and already proven on RUSH. Nothing here changes cohort 1.
The studioRuns a separate Unity lane: the same kit (memory, build farm, publisher desk, review gate) with Unity as a second engine template, tested by the Studios crew with Codex and Claude Code on the studio's own catalogue.
The testTwo weeks, one Windows PC, Unity Pro seats the studio already pays for. Port the RUSH chase loop, or a slice of TAG Police Sentri, to Unity through the agents and ship it to itch.io and Android through the farm. Pass or fail is written down before it starts.

What actually happened with Astra and Unity

15 September. While turning the RUSH browser prototype into a native game, Astra reported that the Android tools and Xcode were present but "Unity isn't installed here", chose Godot for the next installable build, and kept "the production engine decision open until the team confirms the existing Unity project".
The RUSH decisions file records the same: Godot 4.7 for the independent native experiment; production engine open pending the Unity source lineage. The handover lists "a candidate AirAsia Unity repository" among the things to reconcile.
The evidence Astra was looking at. The original Police Sentri shipped on Steam and itch.io in 2021 with Steam integration built for Unity; those integration files are still in a team member's Drive. ToGather: Island is the same generation. Mastra is Unity. So "should we reuse Unity" was the right question for a studio with three Unity titles.

Nothing was decided against Unity. The prototype went Godot because it was the fastest path to an installable build on a Mac with no Unity installed. That is a logistics fact, not an engine verdict.

What Unity offers an AI agent in 2026

Piece	What it is	Why it matters here
Unity AI Assistant (Unity 6.3 LTS)	An agent inside the editor that reads the project and can build scenes, write scripts and generate placeholder assets. Included with Pro, Enterprise and Industry seats. Personal seats pay per use in credits.	The studio's Pro seats already carry it. No new bill to start.
Unity MCP server (official)	A standard bridge that lets outside agents such as Claude Code and Codex operate the editor: create scenes, manage assets, edit scripts, run tasks.	This is how the kit's Build desk reaches Unity the way it reaches Godot today.
Unity's Claude Code plugin (10 Sep 2026)	First-party, distributed through Claude's own plugin directory, bundling 29 Unity skills.	Unity is now maintaining the Claude Code side itself. Less for us to build.
Community bridges (mcp-unity, unity-mcp)	Open-source packages that connect Codex CLI, Cursor and others to the editor; actively maintained in 2026.	Codex, Astra's tool, gets the same door.
AI Gateway	Routes the assistant's work to third-party models such as Claude or GPT; those calls do not spend Unity credits.	Our seats do the thinking; Unity does the editing.
GPT-6 Astra in the wild (from 3 Sep 2026)	Public demos of Astra driving Blender, Godot and Unreal directly, moving menus and building scenes; one showed a full game in about 25 minutes; a round-up thread passed 890,000 views in three days.	This is what Neo pointed Lan to. The public is already doing with Blender and Unreal what we are proposing to do with Unity.

One caution worth stating: Unity's assistant, its credits and its pricing have moved several times in the past two years. The experiment uses the pieces that are stable (the editor, the official server, the Claude Code plugin) and treats the in-editor assistant as a bonus, not a dependency.

What Todak already owns

Unity Pro seats, paid for, with the AI assistant included.
Three Unity titles with source: TAG Police Sentri, ToGather: Island, Mastra. ToGather has sat untouched on Steam for five years and is the obvious first patient.
Steam and itch.io accounts that are live and paying, so a Unity build has somewhere to go on day one (Stores v1).
A crew that knows Unity, and a kit whose backbone, memory, build farm, publisher desk, review gate, does not care which engine produced the build.
RUSH as a yardstick. A chase game with a headless "is it fun" audit, bots and tests. Porting its loop to Unity gives a like-for-like comparison of the agents on two engines.

For Lan: why explore this at all

The studio's catalogue is Unity. An AI pipeline that only speaks Godot cannot touch a single existing Todak game. The value of the pipeline to the studio is measured on its own titles first.
The door is now cheap. Six months ago, making an agent drive Unity meant writing a bridge. Today Unity ships the bridge and the Claude Code plugin itself. The experiment is a fortnight, not a quarter.
The outsourcing pitch depends on it. Todak Studios now sells game development as a service. Clients arrive with Unity projects far more often than Godot ones. "Our AI pipeline works on your engine" is a sentence the studio needs to be able to say.
The public is already there. Neo's point stands: people are sharing Astra building games and Blender scenes unaided. If the studio does not learn how that behaves on its own engine, it learns it from a competitor's showreel.
It does not touch the course. Cohort 1 is Godot. Unity enters the Academy, if ever, as a capstone option for cohort 2, only after this experiment passes.

What Lan is right to worry about. Unity has a history of licence surprises (the 2023 runtime fee, withdrawn in 2024), its AI credits are a metered cost for anyone outside Pro, it is Windows-first, and it is heavier for a beginner than Godot. All four are reasons to keep Unity out of cohort 1, and none of them is a reason for the studio not to test it.

The experiment: two weeks, written pass or fail

Days	Who	What
1–2	Claude Code + one Studios developer	One Windows PC: Unity 6.3 LTS on a Pro seat, the official Unity MCP server, Unity's Claude Code plugin, Codex through the community bridge. Kit's memory and logging hooks attached. Record every prompt, as for a student.
3–6	Codex (Astra) leads, Claude Code reviews	Port the RUSH chase loop to Unity: pasar level, police and robbers, two cameras, touch controls, bots. Astra writes; Claude Code tests and merges, exactly the two-lane rule from the course.
7–8	Claude Code	Port the headless "is it fun" audit (40 rounds, win rates, time fallen) so the Unity build is measured the same way the Godot build was.
9–10	Claude Code + the farm	Android and Windows builds through the build farm, upload to itch.io under the studio account, Play internal testing. Review gate report. Written result.

Pass or fail, decided before day 1

Pass: a playable Unity build of the RUSH loop on an Android phone and on Windows within ten working days; at least 70% of the code and scene work produced by the agents (measured from the prompt record and commits); the build produced by the farm, not by hand; the fun audit runs and returns numbers comparable to the Godot build.
Fail: any of those four missing at day 10. A fail is still useful: it tells the studio exactly where Unity's agent door is narrower than Godot's.
Bonus, not required: a small patch to ToGather: Island shipped to Steam through the same lane, which would also close the housekeeping flag in Stores v1.

What a pass unlocks

Unity becomes the second engine template in the Game Kit, with the same commands.
The studio can quote AI-assisted work on client Unity projects with evidence, not a promise.
Cohort 2 can offer Unity as a capstone option for students who want the industry engine on their CV.
The Roblox lane from Stores v1 runs the same way afterwards: same crew, same fortnight shape, its own engine.

Decisions for Neo & Lan

1Run the two-week Unity experiment with the Studios crew, Codex and Claude Code, on the studio's Pro seats. Recommended
2Which slice: the RUSH chase loop (clean comparison) or TAG Police Sentri's legacy Unity project (reuse test). Neo
3Cohort 1 stays Godot only; Unity enters the Academy as a cohort-2 capstone option only on a pass. Recommended
4Who on the Studios crew pairs with the agents for the fortnight. Neo

Changelog

v1 · 20 Sep 2026. First version, answering Lan's question and Neo's request to let the AI explore the Unity pipeline.