The models got good enough that the harness doesn't matter anymore. So we made the smallest one that proves it. One file. under 200 lines. Zero dependencies.

python3 <(curl -s https://raw.githubusercontent.com/pnegahdar/nano/main/nano.py)

That's the whole thing. No, really.

📂 Reads CLAUDE.md, AGENT.md, AGENTS.md, README.md — automatic repo context, same files the real agents use

🧩 Discovers skill files — finds SKILL.md from .claude/skills, ~/.claude/skills, ~/.codex/skills, ~/.codex/plugins

🛡️ Human-in-the-loop approvals — every command shows what and why before it runs

🔁 200-step agentic loop — keeps going until the job is done, not until it gets bored

💬 Interactive REPL — multi-turn sessions with persistent memory across turns

⚡ Non-interactive CLI — ./nano.py "fix the tests" and walk away

🔄 Session resume — ./nano.py -c picks up where you left off, zero local state (OpenAI stores the conversation)

Plus: interactive REPL · one-shot mode · session resume (-c) · session picker (-s) · GPT-5.5 by default · any model via env var · multi-step tool chaining · approve-one / approve-all / deny · auto-approve mode · per-command cwd, timeout, and env · platform-aware system prompt · output capped at 12KB · forced 5-10 word command descriptions · session reset · zero dependencies · pure stdlib Python · one file you can read in 5 minutes

Most coding agents are built on top of serious infrastructure — frameworks, runtimes, graph orchestrators, thousands of lines of glue.

It turns out you can get surprisingly far with a while loop, one shell tool, and a model that's good enough.

nano.py is that loop. It sends context to GPT-5.5, lets it run shell commands, shows you what it wants to do, waits for your approval, and repeats until the job is done. Pure Python stdlib. Copy it into any repo and go.

# Run it straight from GitHub — no install, no clone, drops into a REPL
python3 <(curl -s https://raw.githubusercontent.com/pnegahdar/nano/main/nano.py)

# Or grab it
curl -O https://raw.githubusercontent.com/pnegahdar/nano/main/nano.py && chmod +x nano.py

# One-shot
OPENAI_API_KEY=sk-... ./nano.py "find the bug in auth.py and fix it"

# Interactive REPL
OPENAI_API_KEY=sk-... ./nano.py

# Continue last session in this directory
OPENAI_API_KEY=sk-... ./nano.py -c

# Pick from recent sessions
OPENAI_API_KEY=sk-... ./nano.py -s

# YOLO mode (auto-approve everything, godspeed)
NANO_APPROVE=all OPENAI_API_KEY=sk-... ./nano.py "run the tests and fix whatever breaks"

• Install packages
• Require a config file
• Need a Docker container
• Have a plugin system

while not done:
    response = ask_model(context)
    if response.wants_to_run_command:
        if human_approves(command):
            result = subprocess.run(command)
            context.append(result)
    else:
        print(response)
        done = True

That's the architecture diagram.

This gives a language model a shell on your computer.

• Approval is on by default. Every command shows you what it wants to run and why.
• You can approve one at a time, or press a to approve all for the session.
• NANO_APPROVE=all skips all prompts. Only use this when you trust the workspace and the task.
• Output is capped at 12KB per command so the model can't filibuster itself.

Read the command before you press y. It's under 200 lines — you can audit the whole thing over lunch.

No score is claimed here. But the shape is right. Same loop every agent runs: read repo, form hypothesis, run command, observe, repeat.

Wrap it in a harness, give it a budget, collect patches, and find out.

Is this a joke?
It works. So no. But also a little bit yes.

Should I use this in production?
Yes. Why not?

Why so small?
The models got good enough that the harness just doesn't need to be big anymore.

Can I swap in Claude / Gemini / etc?
The Responses API is OpenAI-specific, but the pattern is universal. Port it in an afternoon.

Why not just use Claude Code / Codex?
You should. They're great. This is for the mass of developers who want to understand what's actually happening inside the black box — and for anyone who looked at a large agent framework and thought there has to be a simpler way.

There was.

MIT. Copy it, fork it, vendor it into your monorepo.

_{Written by agents, for agents, with agents. The humans just pressed approve.}