On 19 September Unity shipped official plugins for two popular coding agents — Claude Code and Codex. The Codex version carries thirty-one skills: UI, 2D graphics, audio, physics, multiplayer, localisation, in-app purchases. They are written and maintained by Unity's own teams.
The reason is one sentence from the announcement: general-purpose agents lean on forums and tutorials written for older engine versions, and that code compiles but often behaves differently than intended. The agent answers confidently — with the solution from two versions ago.
Where the past comes from
A model knows the world as of its training. Anything newer does not exist for it, and whatever was abundant online is remembered best — even when it is long obsolete. There are always more old articles about the previous version than fresh ones about the current: they had more time to pile up.
Hence the trap: the answer sounds confident and looks plausible, because it used to be correct. The mistake shows up later, when the result has already reached your customer.
What it looks like outside engineering
- The agent quotes a price from the old list — it saw that one more often.
- It recites a policy you changed in June while remembering the March one.
- It cites a regulation in its previous wording.
- It suggests an integration through an API method the vendor has already retired.
- It calculates with a tax rate that expired two years ago.
Note that none of these look like errors. The agent does not hesitate or admit ignorance. It answers in exactly the tone it would use for a correct answer.
How this is fixed
The cure is not a smarter model but where the agent takes its facts from. Four things work, and Unity's solution is essentially the first two.
- Give the agent your sources: price list, policies, database, documentation — so it reads them at answer time instead of recalling them.
- Write your rules down as skills: a short «how we do it here» outweighs any general knowledge the model carries.
- Demand a citation: an answer that cannot name the document it came from is a draft, not an answer.
- Refresh the sources on a schedule and verify the agent sees the new version rather than a forgotten copy.
A ten-minute check
Ask your agent three questions whose answers changed in the last six months: a price, a procedure, some internal rule. Then ask where it got them.
If it cannot name a source, it is answering from the model's memory — and sooner or later it will answer a customer that way. If the source is yours and current, the agent is fine, whatever model sits underneath.
In short
An agent's value comes not from what it knows but from what it is connected to. That is exactly why Unity had to ship its own plugins: giving an agent current knowledge is easier than hoping it forgets the outdated kind.
Source: The Decoder: Unity launches official plugins for Claude Code and OpenAI Codex