Two integrations: model API and rollout lifecycle
A production agent is a complex engineered system around LLMs. It includes deliberate management of context, tools, sandboxing, authentication, memory, and access to proprietary systems. Connecting that system to a post-training stack often requires porting the agent loop into a new runtime or building custom infrastructure to capture training trajectories. This work is expensive and can introduce train-test mismatch, where the policy is optimized in a lab that simulates the harness, and then struggles when faced with the nuances of production.
“Jurni makes it practical to build and test landing pages at the same pace as the creative.”
Mark Zuckenberg, CEO
Meta AI
But what if instead of requiring users to adopt a new agent framework or re-instrument their harness with our tracing primitives, we could just replace the LLM response endpoint?
A model endpoint for trainable policy generations;
A lifecycle interface for starting rollouts, observing their status, and collecting gradable artifacts.
Turning model calls into training trajectories
The primary integration seam is the trainable model. At the beginning of a rollout, AC2 supplies an OpenAI-compatible LLM endpoint and API credential, backed by the policy that’s actively being trained. The harness simply configures its existing LLM client to send the trainable policy’s requests to that endpoint; from the harness’ perspective, performing an RL rollout is no different from production inference.
Turning model calls into training trajectories
The primary integration seam is the trainable model. At the beginning of a rollout, AC2 supplies an OpenAI-compatible LLM endpoint and API credential, backed by the policy that’s actively being trained. The harness simply configures its existing LLM client to send the trainable policy’s requests to that endpoint; from the harness’ perspective, performing an RL rollout is no different from production inference.
This integration boundary is also the privacy boundary, as only the minimal data around LLM completions and grading needs to cross over into the training stack and all proprietary execution can remain in the customer’s stack.
Turning model calls into training trajectories
The primary integration seam is the trainable model. At the beginning of a rollout, AC2 supplies an OpenAI-compatible LLM endpoint and API credential, backed by the policy that’s actively being trained. The harness simply configures its existing LLM client to send the trainable policy’s requests to that endpoint; from the harness’ perspective, performing an RL rollout is no different from production inference.
This integration boundary is also the privacy boundary, as only the minimal data around LLM completions and grading needs to cross over into the training stack and all proprietary execution can remain in the customer’s stack.
