Nano

A minimal coding‑agent for:

1. agent‑in‑the‑loop reinforcement learning
   
2. understanding coding agents in clear, minimal terms
   
3. running neat little code fixes

---

What it is

Nano is a zero‑bloat wrapper that turns any tool-enabled LLM into a coding agent with two tools:

```

shell(cmd) # ls, cat, grep … apply_patch({...}) # search/replace on one file

```

Note: Nano uses rbash (restricted bash) to confine the agent to its designated workspace. This, along with Nano's requirement of starting in a clean git repository, helps ensure its operations remain contained and predictable.

---

Why it exists

Most coding agents (e.g. Aider, SWE-Agent, Devin) are designed to perform well. To achieve that, they bake in layers of effective but ad-hoc solutions:
repo maps, navigation memory, multi agent orchestration, adherence prompting, retry logic,...

These make agents more capable, but also more opaque. They're hard to analyze, and thus harder to adopt.

Nano takes the opposite stance:

Inspired by The Bitter Lesson, we believe that long-term performance comes not from encoding human intuitions, but from letting models learn their own strategies, even if they start out worse.

Effective reinforcement learning relies on a complete and unaltered log of agent interactions. Nano ensures this transparency by providing direct, non-obfuscated access to the raw reasoning, tool calls, and results, offering a precise record of what the model saw and did.

That's what Nano tries to provide.

---

Install

```bash git clone git@github.com:ASSERT-KTH/nano-agent.git && cd nano-agent && pip install -e .

or

pip install nano-agent ```

Then you just need an API key for your chosen provider or host them yourself with vLLM. See litellm documentation for more details.

---

Example: rollout to Tensor

```python from nano import Agent from transformers import AutoTokenizer

agent = Agent(model="openai/gpt-4.1-mini") agent.run("There is a bug in this repo...")

tokenizer = AutoTokenizer.frompretrained("Qwen/Qwen3-8B") tokens = tokenizer.applychattemplate( agent.messages, tools=agent.tools, tokenize=True, returnformat="pt" ) ```

Example: minimal SWE‑Gym rollout

```python import tempfile from git import Repo # git-python from nano import Agent from datasets import load_dataset

run = load_dataset("SWE-Gym/SWE-Gym", split="train[:1]")[0]

tempdir = tempfile.mkdtemp() Repo.clone_from(f"https://github.com/{run['repo']}.git", tempdir)

agent = Agent( model="hostedvllm/qwen/qwen3-8b", apibase="http://localhost:8000/v1", ) diff = agent.run(run["problemstatement"], reporoot=tempdir) print(diff) # the unified diff produced by the agent print(agent.messages, agent.tools) # or access in ~/.nano/<timestamp>/ ``

---

Use with HuggingFace TRL

Because Nano can communicate with any tool-enabled OpenAI compatible endpoint and produces token-level message logs, it works "cleanly" as a data generator inside TRL's GPROTrainer.

Note: "cleanly" refers to modifications made in our TRL fork to enable direct agent integration. These changes support the CodeRepairRL project but may not be merged into the main HuggingFace repository.

To use it:

• Write a rollout client that wraps Agent.run()
• Extract the diff and messages for each training example
• Feed those into TRL's reward modeling or fine-tuning pipelines

This lets you train models that learn to use tools directly, grounded in interaction data — no custom env needed.

This approach acknowledges that the agent may initially fail in certain situations; however, these failures are valuable learning opportunities. We can then directly reinforce favorable behaviors and successful outcomes using outcome supervision, progressively refining the agent's strategies.

---

Citation

@misc{nano-agent2025, author = {Bjarni Haukur}, title = {Nano: a minimalist coding agent for agent-in-the-loop training}, howpublished = {\url{https://github.com/ASSERT-KTH/nano-agent}}, year = {2025} }

🏆 Current Leaderboard

Performance on SWE-bench Lite subset, ranked by code similarity

#	Ver	Model	Code Sim	Test Sim	Tokens	Tools
1	v3.2.0	claude-sonnet-4-20250514	0.394	0.188	14,746 / 16,384	41.5 / 100
2	v3.2.0	gpt-4.1	0.387	0.092	9,777 / 16,384	35.7 / 100
3	v4.0.1	kimi-k2	0.382	0.009	5,508 / 16,384	19.7 / 100
4	v4.0.1	qwen3-coder	0.374	0.042	6,979 / 16,384	26.5 / 100
5	v3.2.0	gemini-2.5-pro-preview	0.370	0.034	6,008 / 16,384	13.6 / 100
6	v3.3.0	gemini-2.5-flash	0.363	0.022	4,337 / 16,384	13.2 / 100
7	v3.2.0	gemini-2.5-flash-preview-05-20	0.362	0.000	4,547 / 16,384	10.1 / 100
8	v3.2.0	gpt-4.1-mini	0.350	0.017	7,403 / 16,384	29.7 / 100
9	v3.2.0	deepseek-chat	0.336	0.011	3,297 / 16,384	7.5 / 100
10	v4.0.1	glm-4.5	0.323	0.107	12,477 / 16,384	28.7 / 100
11	v3.2.0	qwen-2.5-72b-instruct	0.272	0.000	5,873 / 16,384	35.1 / 100
12	v3.2.0	qwen3-32b	0.255	0.000	5,281 / 16,384	28.3 / 100
13	v3.2.0	llama-4-maverick	0.255	0.000	4,647 / 16,384	10.4 / 100
14	v3.2.0	qwen3-14b-thinking	0.253	0.000	8,549 / 16,384	16.3 / 100
15	v3.3.0	gemini-2.5-flash-lite-preview-06-17	0.243	0.005	6,294 / 16,384	21.6 / 100
16	v3.2.0	qwen3-32b-thinking	0.224	0.005	9,357 / 16,384	8.3 / 100
17	v3.2.0	qwen3-8b-thinking	0.210	0.000	8,688 / 16,384	15.0 / 100
18	v3.2.0	qwen3-8b	0.190	0.000	8,704 / 16,384	56.5 / 100
19	v3.2.0	gpt-4.1-nano	0.188	0.000	8,536 / 16,384	33.1 / 100
20	v3.2.0	qwen3-14b	0.176	0.000	10,800 / 16,384	82.6 / 100
21	v3.2.0	devstral-small	0.092	0.000	14,603 / 16,384	13.0 / 100

🏆 SWE-bench Verified Leaderboard

Performance on SWE-bench Verified subset, ranked by code similarity

#	Ver	Model	Code Sim	Test Sim	Tokens	Tools
1	v3.3.2	gemini-2.5-flash	0.353	0.000	5,956 / 16,384	12.5 / 100

How it works: - Input: A GitHub repository containing a bug with a known ground truth solution - Task: Nano provides models with tools to explore the codebase and generate a fix - Output: Nano produces a unified git diff containing all proposed code changes - Evaluation: We measure how closely the model's solution matches the ground truth using: - Code Similarity: How well the fix matches the actual bug fix (primary ranking metric) - Test Similarity: How well any test changes match the ground truth test updates

Note: Prone to a lot of noise, small test set with few repetitions.