[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:mini-swe-agent":3},"\u003Cdiv align=\"center\">\n\u003Ca href=\"https:\u002F\u002Fmini-swe-agent.com\u002Flatest\u002F\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002Fmini-swe-agent\u002Fraw\u002Fmain\u002Fdocs\u002Fassets\u002Fmini-swe-agent-banner.svg\" alt=\"mini-swe-agent banner\" \u002F>\u003C\u002Fa>\n\u003C\u002Fdiv>\u003Ch1>The minimal AI software engineering agent\u003C\u002Fh1>\n\u003Cp>📣 \u003Ca href=\"https:\u002F\u002Flabs.ramp.com\u002Fswebench\" rel=\"nofollow ugc noopener\">mini-swe-agent now powers Ramp SWE-Bench\u003C\u002Fa>\u003Cbr \u002F>\n📣 \u003Ca href=\"https:\u002F\u002Fdeepswe.datacurve.ai\u002Fblog#evaluation-harness\" rel=\"nofollow ugc noopener\">mini-swe-agent beats Claude Code and Codex on DeepSWE\u003C\u002Fa>\u003Cbr \u002F>\n📣 \u003Ca href=\"https:\u002F\u002Fmini-swe-agent.com\u002Flatest\u002Fusage\u002Fprogrambench\u002F\" rel=\"nofollow ugc noopener\">Run mini-swe-agent on our new & extremely challenging benchmark, ProgramBench\u003C\u002Fa>\u003Cbr \u002F>\n📣 \u003Ca href=\"https:\u002F\u002Fminimal-agent.com\u002F\" rel=\"nofollow ugc noopener\">New tutorial on building minimal AI agents\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Ca href=\"https:\u002F\u002Fmini-swe-agent.com\u002Flatest\u002F\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocs-green?style=for-the-badge&logo=materialformkdocs&logoColor=white\" alt=\"Docs\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fjoin.slack.com\u002Ft\u002Fswe-bench\u002Fshared_invite\u002Fzt-36pj9bu5s-o3_yXPZbaH2wVnxnss1EkQ\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FSlack-4A154B?style=for-the-badge&logo=slack&logoColor=white\" alt=\"Slack\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fpypi.org\u002Fproject\u002Fmini-swe-agent\u002F\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Fmini-swe-agent?style=for-the-badge&logo=python&logoColor=white&labelColor=black&color=deeppink\" alt=\"PyPI - Version\" \u002F>\u003C\u002Fa>\u003C\u002Fp>\n\u003Cblockquote>\n\u003Cp>[!WARNING]\nThis is \u003Cstrong>mini-swe-agent v2\u003C\u002Fstrong>. Read the \u003Ca href=\"https:\u002F\u002Fmini-swe-agent.com\u002Flatest\u002Fadvanced\u002Fv2_migration\u002F\" rel=\"nofollow ugc noopener\">migration guide\u003C\u002Fa>. For the previous version, check out the \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002Fmini-swe-agent\u002Ftree\u002Fv1\" rel=\"nofollow ugc noopener\">v1 branch\u003C\u002Fa>.\u003C\u002Fp>\n\u003C\u002Fblockquote>\n\u003Cp>In 2024, we built \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fswe-bench\u002FSWE-bench\" rel=\"nofollow ugc noopener\">SWE-bench\u003C\u002Fa> & \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fswe-agent\u002Fswe-agent\" rel=\"nofollow ugc noopener\">SWE-agent\u003C\u002Fa> and helped kickstart the coding agent revolution.\u003C\u002Fp>\n\u003Cp>We now ask: \u003Cstrong>What if our agent was 100x simpler, and still worked nearly as well?\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>\u003Ccode>mini\u003C\u002Fcode> is\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Widely adopted\u003C\u002Fstrong>: Used by Meta, NVIDIA, Essential AI, IBM, Nebius, Anyscale, Princeton University, Stanford University, and many more.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Minimal\u003C\u002Fstrong>: Just some 100 lines of python for the \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002Fmini-swe-agent\u002Fblob\u002Fmain\u002Fsrc\u002Fminisweagent\u002Fagents\u002Fdefault.py\" rel=\"nofollow ugc noopener\">agent class\u003C\u002Fa> (and a bit more for the \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002Fmini-swe-agent\u002Fblob\u002Fmain\u002Fsrc\u002Fminisweagent\u002Fenvironments\u002Flocal.py\" rel=\"nofollow ugc noopener\">environment\u003C\u002Fa>,\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002Fmini-swe-agent\u002Fblob\u002Fmain\u002Fsrc\u002Fminisweagent\u002Fmodels\u002Flitellm_model.py\" rel=\"nofollow ugc noopener\">model\u003C\u002Fa>, and \u003Ca href=\"https:\u002F\u002Fgithub.com\u002FSWE-agent\u002Fmini-swe-agent\u002Fblob\u002Fmain\u002Fsrc\u002Fminisweagent\u002Frun\u002Fhello_world.py\" rel=\"nofollow ugc noopener\">run script\u003C\u002Fa>) — no fancy dependencies!\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Performant:\u003C\u002Fstrong> Scores >74% on the \u003Ca href=\"https:\u002F\u002Fwww.swebench.com\u002F\" rel=\"nofollow ugc noopener\">SWE-bench verified benchmark\u003C\u002Fa>; starts much faster than Claude Code\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Deployable:\u003C\u002Fstrong> Supports \u003Cstrong>local environments\u003C\u002Fstrong>, \u003Cstrong>docker\u002Fpodman\u003C\u002Fstrong>, \u003Cstrong>singularity\u002Fapptainer\u003C\u002Fstrong>, \u003Cstrong>bublewrap\u003C\u002Fstrong>, \u003Cstrong>contree\u003C\u002Fstrong>, and more\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Compatible:\u003C\u002Fstrong> Supports all models via \u003Cstrong>litellm\u003C\u002Fstrong>, \u003Cstrong>openrouter\u003C\u002Fstrong>, \u003Cstrong>portkey\u003C\u002Fstrong>, and more. Support for \u003Ccode>\u002Fcompletion\u003C\u002Fcode> and \u003Ccode>\u002Fresponse\u003C\u002Fcode> endpoints, interleaved thinking etc.\u003C\u002Fli>\n\u003Cli>Built by the Princeton & Stanford team behind \u003Ca href=\"https:\u002F\u002Fswebench.com\" rel=\"nofollow ugc noopener\">SWE-bench\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fswe-agent.com\" rel=\"nofollow ugc noopener\">SWE-agent\u003C\u002Fa>, and more\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Tested:\u003C\u002Fstrong> \u003Ca href=\"https:\u002F\u002Fcodecov.io\u002Fgh\u002FSWE-agent\u002Fmini-swe-agent\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fcodecov\u002Fc\u002Fgithub\u002Fswe-agent\u002Fmini-swe-agent?style=flat-square\" alt=\"Codecov\" \u002F>\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cdetails>\u003Csummary>More motivation (for research)\u003C\u002Fsummary>\u003Cp>\u003Ca href=\"https:\u002F\u002Fswe-agent.com\u002Flatest\u002F\" rel=\"nofollow ugc noopener\">SWE-agent\u003C\u002Fa> 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.\nHowever, one year later, as LMs have become more capable, a lot of this is not needed at all to build a useful agent!\nIn fact, the \u003Ccode>mini\u003C\u002Fcode> agent\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Does not have any tools other than bash\u003C\u002Fstrong> — it doesn't even need to use the tool-calling interface of the LMs.\nThis means that you can run it with literally any model. When running in sandboxed environments you also don't need to take care\nof installing a single package — all it needs is bash.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Has a completely linear history\u003C\u002Fstrong> — every step of the agent just appends to the messages and that's it.\nSo there's no difference between the trajectory and the messages that you pa\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fdetails>",1784853423792]