SWE-agent 1.0.0 migration guide

Welcome to SWE-agent 1.0! So much new stuff! Here's a quick rundown of the cool new things you can do:

• :sparkles: Fast, massively parallel code execution with SWE-ReX.
• :sparkles: Run SWE-agent locally but execute code in the cloud (using modal, AWS, or anything else that runs SWE-ReX).
• :sparkles: Configurable retry mechanisms: Try multiple agent configurations, models, parameters, etc., then choose the best one.
• :sparkles: Flexible tool definitions with tool bundles.
• :sparkles: All language models supported using litellm (see models).
• :sparkles: Override any configuration option from the command line (see command line basics).
• :sparkles: New command line trajectory inspector to scroll few hundreds of trajectories with ease.
• :sparkles: New command line interface with subcommands for running over single issues, batches, and various utility commands.
• :sparkles: Greatly simplified and cleaned up codebase. In particular, the Agent class is now much easier to modify.

If you're familiar with the old SWE-agent, here are the main changes you need to be aware of.

!!! tip "SWE-agent 0.7" The old SWE-agent before the major overhaul is still preserved on the v0.7 branch. In particular, this is still the version to use for the SWE-agent EnIGMA configuration. You can also switch to the old documentation using the dropdown in the top menu bar.

Command line interface and basic configuration

• Instead of the run.py script, we now have a central sweagent entry point that you can call after installing the package. This should be much more convenient!
• Instead of one run.py command, we now have several subcommands: sweagent run to run over single issues, sweagent run-batch to run over a whole batch of issues, and various utility commands. Run sweagent --help to see all options. Splitting up both commands made it easier to make both use cases more convenient and flexible.
• We have switched to a hierarchical configuration system. This means that command line options look something like this: --agent.model.name=gpt-4o. The configuration files have also been updated to reflect this.

Environment setup

We removed the complicated mess of environment setup options (conda environments, pip, docker images, etc.). Instead, you now always start from a docker image of your choice and we recommend that this should ship with all the dependencies you need. However, you can also execute additional commands before starting the agent with EnvironmentConfig.post_startup_commands. Additionally, every tool bundle can include a setup.sh script that will be executed, allowing to e.g., install flake8 if needed by the tools.

More advanced configuration

• Tools now live in tool bundles in the tools/ directory. You can mix and match tools to your liking.

Code

The codebase has been nearly rewritten from scratch and both more powerful and more flexible.

• The biggest change is SWE-ReX, our new "backend" that handles all code execution.
• As a result of this, the SWEEnv class is basically gone and only is a small wrapper around a swerex runtime
• The Agent class also has gotten a lot simpler. It also delegates a lot of tool/execution logic to the new Tools class.