Definition
Plain language
A small trained AI that sits between an agent and the outside world, deciding which past information reaches the model and which of the model's proposed actions are allowed through.
As stated in the literature
A learnable bidirectional controller for LLM agent harnesses, fine-tuned via distillation to perform observation projection (pass, compress, or drop each past turn while keeping the full record intact) and action projection (evidence-grounded gatekeeping of proposed tool calls); trained on one model's traces and transferred zero-shot to other models, since it manipulates only the conversation, never the weights.
Why it matters: It controls what an agent sees and does without retraining the underlying model, and can be reused across different models without starting over.
For example, before a question reaches the model it might compress ten earlier turns into one summary, and before the model's chosen action runs it might block one that lacks supporting evidence.
Heard on the show
“It's called "HarnessBridge: Learnable Bidirectional Controller for LLM Agent Harness," and what you're hearing is an AI-generated show — the script was written by Anthropic's Claude Opus 4.”Episode 142 — Training a Tiny Model to Run the Plumbing Between an Agent and the World