Four training frameworks adopt Agent Lightning's proxy pattern

Four training frameworks now route agent traffic through one shared proxy, per a paper on Agent Lightning's design. verl Uni-Agent, AReaL 2.0, slime and Polar all watch their agents the same way Agent Lightning does, by sitting an LLM request/response proxy between the agent and the model instead of instrumenting the agent's own code.

That interception point turns out to be useful for more than training. Spend caps against runaway agent costs live there too, per the same source's resource-hijacking findings. So does token accounting. Model A/B testing has been drifting toward the same seam.

Four separate problems landing on one integration point is usually a sign the point itself is the real boundary, not an implementation detail specific to any one of them.

The take

If you're running agents against a model provider without a proxy in front of that traffic, you're going to build one eventually. The order matters: doing it before you have three separate orchestrators each talking to the provider directly is cheaper than retrofitting a proxy under all three later.

The paper doesn't say whether these four frameworks converged on the design independently or copied Agent Lightning outright, a gap worth watching as more training stacks publish their own architectures.