The 100 line AI agent that's actually useful

This is mini-swe-agent v2

Read the migration guide. For the previous version, check out the v1 documentation or the v1 branch.

In 2024, SWE-bench & SWE-agent helped kickstart the coding agent revolution.

We now ask: What if our agent was 100x simpler, and still worked nearly as well?

mini is

Widely adopted: Used by Meta, NVIDIA, Essential AI, IBM, Nebius, Anyscale, Princeton University, Stanford University, and many more.
Minimal: Just 100 lines of python (+100 total for env, model, script) — no fancy dependencies!
Performant: Scores >74% on the SWE-bench verified benchmark; starts much faster than Claude Code
Deployable: Supports local environments, docker/podman, singularity/apptainer, bublewrap, contree, and more
Compatible: Supports all models via litellm, openrouter, portkey, and more. Support for /completion and /response endpoints, interleaved thinking etc.
Built by the Princeton & Stanford team behind SWE-bench, SWE-agent, and more
Tested:

Why use mini-SWE-agent for research?

SWE-agent jump-started the development of AI agents in 2024. Back then, we placed a lot of emphasis on tools and special interfaces for the agent. However, one year later, a lot of this is not needed at all to build a useful agent!

In fact, the mini agent:

Does not have any tools other than bash — it doesn't even use the tool-calling interface of the LMs. This means that you can run it with literally any model. When running in sandboxed environments you also don't need to take care of installing a single package — all it needs is bash.
Has a completely linear history — every step of the agent just appends to the messages and that's it. So there's no difference between the trajectory and the messages that you pass on to the LM. Great for debugging & fine-tuning.
Executes actions with subprocess.run — every action is completely independent (as opposed to keeping a stateful shell session running). This makes it trivial to execute the actions in sandboxes (literally just switch out subprocess.run with docker exec) and to scale up effortlessly. Seriously, this is a big deal, trust me.

This makes it perfect as a baseline system and for a system that puts the language model (rather than the agent scaffold) in the middle of our attention. You can see the result on the SWE-bench (bash only) leaderboard, that evaluates the performance of different LMs with mini.

Why use mini-SWE-agent as a tool?

Some agents are overfitted research artifacts. Others are UI-heavy frontend monsters.

The mini agent wants to be a hackable tool, not a black box.

Simple enough to understand at a glance
Convenient enough to use in daily workflows
Flexible to extend

Unlike other agents (including our own swe-agent), it is radically simpler, because it:

Does not have any tools other than bash — it doesn't even use the tool-calling interface of the LMs. Instead of implementing custom tools for every specific thing the agent might want to do, the focus is fully on the LM utilizing the shell to its full potential. Want it to do something specific like opening a PR? Just tell the LM to figure it out rather than spending time to implement it in the agent.
Executes actions with subprocess.run — every action is completely independent (as opposed to keeping a stateful shell session running). This is a big deal for the stability of the agent, trust me.
Has a completely linear history — every step of the agent just appends to the messages that are passed to the LM in the next step and that's it. This is great for debugging and understanding what the LM is prompted with.

Should I use mini-SWE-agent or swe-agent?

You should consider mini-swe-agent your default choice. In particular, you should use mini-swe-agent if

You want a quick command line tool that works locally
You want an agent with a very simple control flow
You want even faster, simpler & more stable sandboxing & benchmark evaluations
You are doing FT or RL and don't want to overfit to a specific agent scaffold

You should use swe-agent if

You want to experiment with different sets of tools, each with their own interface
You want to experiment with different history processors

What you get with both

Excellent performance on SWE-Bench
A trajectory browser

CLI (`mini`)	Batch inference

Trajectory browser	Python bindings
	`agent = DefaultAgent( LitellmModel(model_name=...), LocalEnvironment(), ) agent.run("Write a sudoku game")`