---
license: apache-2.0
base_model: Qwen/Qwen3.6-27B
library_name: transformers
language:
  - en
  - zh
tags:
  - agents
  - agentic-systems
  - harness-generation
  - tool-use
  - code-generation
  - jit-agent
---

# JIT-Agent-27B

**JIT-Agent-27B** is a harness intelligence model for synthesizing executable,
task-conditioned agent harnesses. Given a task, an available tool registry, a
shared runtime protocol, and natural-language descriptions of reference
harnesses, the model generates a complete operational scaffold tailored to the
task at hand.

Instead of directly solving the task, JIT-Agent writes the system through which
another foundation model acts: how it maintains memory, forms and updates plans,
executes actions, and orchestrates tools and skills.

## Checkpoint Overview

This repository contains the initial research release of JIT-Agent-27B. The
checkpoint builds on the Stage-I harness-customization model and is further
trained through distillation from the final research checkpoint.

## Model Details

| Property | Value |
|---|---|
| Base model | [`Qwen/Qwen3.6-27B`](https://huggingface.co/Qwen/Qwen3.6-27B) |
| Parameters | 27.36B |
| Weight precision | BF16 |
| Architecture context length | 262,144 tokens |
| Recommended serving context | 163,840 tokens |
| Primary input | Task, tools, protocol, and reference-harness descriptions |
| Primary output | Four Python modules and one YAML prompt configuration |

JIT-Agent generates harnesses under a fixed four-module protocol:

- `memory.py`: constructs and updates the agent's working context;
- `planning.py`: forms directives and manages plan state;
- `action.py`: implements the task-execution loop;
- `tool_policy.py`: controls tool and skill exposure;
- `prompt.yaml`: defines the prompts consumed by the generated modules.

The model emits these files using the following tagged format:

```text
<<<PYTHON_MEMORY>>>
...
<<<END_PYTHON_MEMORY>>>

<<<PYTHON_PLANNING>>>
...
<<<END_PYTHON_PLANNING>>>

<<<PYTHON_ACTION>>>
...
<<<END_PYTHON_ACTION>>>

<<<PYTHON_TOOL_POLICY>>>
...
<<<END_PYTHON_TOOL_POLICY>>>

<<<YAML>>>
...
<<<END_YAML>>>
```

## Recommended Usage

The checkpoint is designed to be used with the
[JIT-Agent runtime](https://github.com/bingreeky/JIT), which constructs the full
generation prompt, validates the structured output, installs the resulting
harness, and executes it against an off-the-shelf agentic model.

### 1. Set up the runtime

```bash
git clone https://github.com/bingreeky/JIT.git
cd JIT

conda env create -f environment.yml
conda activate jit
```

### 2. Serve the checkpoint

The repository provides a vLLM launcher with the recommended serving settings:

```bash
MODEL=JIT-Agent/jit-27b \
TP=4 \
bash scripts/serve_meta_model.sh
```

This exposes an OpenAI-compatible endpoint at `http://localhost:8000/v1`.

### 3. Generate and execute a task-conditioned harness

```bash
python -m scripts.run_jit \
    --bench xbench \
    --meta-model jit \
    --meta-base http://localhost:8000/v1 \
    --harness-refs desc \
    --max-samples 5
```

The released checkpoint should be used with description references:

```text
--harness-refs desc
```

In this mode, the model receives natural-language design descriptions of the
reference harnesses rather than their source code. This is also the default mode
of the released runtime.

For best-of-N inference, the runtime generates three candidate harnesses at
temperature 1.0 and selects one before task execution.

## Intended Use

JIT-Agent-27B is intended for research on:

- task-adaptive agent architecture generation;
- model–harness co-design;
- modular agent runtimes;
- memory, planning, action, and tool-policy composition;
- best-of-N harness synthesis;
- harness transfer across executor models and task domains.

It is a harness generator rather than a general-purpose chat model. Direct
chat-style prompting without the accompanying protocol and runtime context is
unlikely to produce valid executable harnesses. Also see [here](https://huggingface.co/papers/2608.25593).

## License

The checkpoint is released under the Apache License 2.0. It is derived from
[`Qwen/Qwen3.6-27B`](https://huggingface.co/Qwen/Qwen3.6-27B), which is also
released under Apache 2.0.