For coding agents, adding more tools does not automatically make the system more capable. This project takes the counterintuitive route: keep the agent scaffold small and let the language model use Bash directly. That simplicity makes it easier to inspect what the model sees, reproduce its actions, and compare models without tying results to a large agent framework.
What Sets It Apart
- Bash-only interaction: The model works through the shell rather than a collection of custom tools, so the same agent can work with a wide range of models and environments.
- Independent command execution: Actions run separately through subprocesses instead of relying on a persistent shell session. This makes sandboxing and parallel evaluation simpler.
- Linear interaction history: Each step is appended to the conversation, making the trajectory easier to follow and useful for debugging or fine-tuning. The project reports over 74% on SWE-bench Verified.
Who It Suits—and the Trade-off
A good fit for developers who want a compact local coding agent, or researchers who need a stable baseline for benchmark evaluation, fine-tuning, or reinforcement learning. Look elsewhere if you want to experiment with many specialized tools or custom history processing; the deliberately minimal interface leaves less room for those extensions than SWE-agent.