bingreeky/JIT

What it solves

JIT-Agent addresses the limitation of using a single, general-purpose agent scaffold (harness) for all tasks. Instead of relying on one static design, it enables the creation of task-specific harnesses that are optimized for the particular requirements of a given problem, improving the performance of off-the-shelf agentic LLMs.

How it works

JIT-Agent acts as a meta-agent that generates executable, task-specific harnesses on the fly. It takes a task specification, a protocol, a tool registry, and prior harness examples to emit structured code. Each harness is composed of four modular components: memory, planning, action, and capability orchestration. These modules are implemented against shared interfaces in a "HarnessFactory". As the agent receives feedback and traces, the meta-agent revises the harness to improve performance at test time, while the generator itself remains frozen.

Who it’s for

It is designed for developers and researchers working with agentic LLMs who want to maximize the effectiveness of their agents across diverse benchmarks (such as deep research, planning, and workspace tasks) without manually redesigning the scaffold for every new task.

Highlights

• Model-as-a-Harness: Wraps any off-the-shelf LLM in a dynamically generated, task-specific scaffold.
• Modular Architecture: Factors harnesses into four distinct modules (memory, planning, action, and capability).
• Just-in-Time Evolution: Harnesses are revised and improved based on feedback during the testing phase.
• Transferable Intelligence: Treats the agent scaffold as a trainable, transferable axis of intelligence separate from the base model scaling.