27B open-source JIT-Agent writes task-specific harnesses on the fly instead of one fixed scaffold

blaizedsouza · x · 2026-09-04

The open-source JIT-Agent project flips the fixed-harness model: a 27B meta-model takes a task spec, protocol, tool/skill registry and retrieved prior harnesses, and emits an executable, task-specific harness wrapping any off-the-shelf agentic LLM (Model-as-a-Harness).
Harness generation happens once before the task: four Python files plus a prompt config following a fixed four-module contract — memory controls what history slice the worker model sees, planning turns that view into the next directive, tool policy decides which tools are exposed.
The second model doing the actual work is never trained, never writes code, and never sees the harness files; it is just called step by step. The author also published a deep-dive article on the anatomy of agent harnesses built by Anthropic, OpenAI, Perplexity and LangChain.

More from coding & agent