GitHub's Java agent runtime ships as a Maven dependency, and the tool schema comes from reflection

The load-bearing detail in this design is reflection. You register a tool by passing a method reference to `session.registerTool()`, and the SDK reads parameter types, names and the return type off the signature to emit an OpenAI function definition ^[7]. That is less ceremony than a `FunctionCallback` bean or a `@Tool` annotation ^[8]. It also moves the tool contract from something you declare to something your build emits: the names the model sees are the names the compiled artifact retained ^[13]. Anyone evaluating this should check what their compiler flags and any bytecode processing in the pipeline do to parameter names before checking anything else.

The sample controller in the writeup shows where the rest of the cost sits. The client is built in the constructor with an environment variable for the key and the string `gpt-4` for the model, and both the session and the two tool registrations happen inside the request handler ^[11]. Multi-model routing that lets you switch models without changing application code is on the capability list ^[6]; the example changes application code to pick a model ^[11]. Per-request registration also means the reflection and schema build sit on the request path unless the SDK caches them, which the writeup does not address.

None of this has been through anyone's production. The author states it directly: the evaluation rests on GitHub's official documentation, the SDK README and engineering posts from the release team, and he has not shipped the SDK himself ^[10]. So the "production-tested" credit is inherited from Copilot CLI's lineage as the same embedded runtime ^[2], not from a report of it running in a servlet container. The capability list is a specification. Read it as one.

What the third path actually buys is narrower than the framing suggests. Adopting Spring AI or LangChain4j means adopting their abstractions, their release cadence and their opinion of what an agent loop is ^[9]. The Copilot SDK does not remove that dependency; it relocates it into a compiled runtime you cannot see into, and in BYOK mode the loop runs in your JVM, which the author correctly identifies as the mode where you own rate limits, observability and failure modes ^[5]. That is the trade in one sentence: you stop writing to someone's abstraction and start operating someone's control flow.

The portability arithmetic is worth stating plainly, because none of the three options talk to each other. Spring AI wants beans, LangChain4j wants annotations, the Copilot SDK wants method references: three distinct registration mechanisms, so a tool layer written for any one of them is rewritten on any move ^[12]. For a service with a handful of tools that is an afternoon. For a service with forty, the choice made this quarter on the strength of a docs-based writeup becomes the thing nobody wants to revisit next year.