Mini-SWE-agent

Mini-SWE-agent platform architecture and minimalist agent loop

Mini-SWE-agent (developed by the research team at Princeton NLP & Stanford University, accessible via mini-swe-agent.com and GitHub) is an open-source, ultra-lightweight autonomous software engineering agent. Designed as the radical evolution of the original SWE-agent framework that helped ignite the coding agent revolution in 2024, Mini-SWE-agent proves that an AI agent implemented in just 100 lines of clean Python can match and exceed the performance of monolithic, tool-heavy agent frameworks.

Distributed under the permissive MIT license, Mini-SWE-agent strips away bloated scaffolds, stateful shell watchers, and complex tool-calling abstractions. Instead, it places the reasoning capabilities of modern language models front and center, executing tasks purely through standard bash commands and strict linear conversation histories. Adopted globally by research labs and enterprises including Meta, NVIDIA, Essential AI, IBM, Nebius, and Anyscale, Mini-SWE-agent provides the gold standard for reproducible agent evaluations.

Pure bash execution, linear history, and pluggable sandboxes

Mini-SWE-agent delivers exceptional architectural clarity and engineering advantages:

• Radical 100-Line Codebase: The entire core agent loop is implemented in approximately 100 lines of standard Python code, making it instantly readable, fully auditable, and trivial to customize for bespoke research and fine-tuning.
• Pure Bash Tool Interface: Does not rely on vendor-locked tool-calling schemas or custom file-editing commands. The model navigates codebases, performs edits, runs compilers, and executes test suites using standard shell tools (sed, awk, pytest, git).
• Stateless Subprocess Execution: Runs each command independently via subprocess.run or container execution calls rather than maintaining fragile stateful shell sessions, eliminating hang-ups and desynchronization bugs.
• Strict Linear History: Eliminates hidden prompt injections and complex history processors. The agent trajectory maps 1:1 to the message array passed to the model, making debugging and trajectory inspection completely transparent.
• Universal Model Support via LiteLLM: Seamlessly connects to leading frontier models including Anthropic Claude Opus 5, Claude Sonnet 5, OpenAI GPT-5.3-Codex, Google Gemini 3.7 Flash, DeepSeek-V3, as well as local GGUF/vLLM inference engines.
• Pluggable Sandboxing Backends: Natively executes inside local environments, Docker containers, Podman, Singularity/Apptainer, Bubblewrap, and Contree for high-throughput batch evaluation.

Comparative benchmark: Mini-SWE-agent vs. SWE-agent and OpenHands

Mini-SWE-agent demonstrates that minimalist agent scaffolding delivers higher stability and faster execution on standardized benchmarks.

Research workflows, SWE-bench evaluations, and operational guardrails

• Automated benchmark evaluation: Execute large-scale batch runs on SWE-bench Verified, Ramp SWE-bench, and ProgramBench to benchmark open and proprietary model capabilities.
• Fine-tuning and reinforcement learning (RL): Collect clean, uncorrupted agent trajectories without scaffolding artifacts to train specialized code-generation models.
• Local CLI bug fixing: Run the interactive mini command line tool inside local repositories to autonomously resolve GitHub issues and unit test failures.
• Visual trajectory inspection: Launch the built-in web-based trajectory inspector to review step-by-step model reasoning, commands, and bash outputs.

Operational guardrails: Mini-SWE-agent runs by default inside isolated Docker or container sandboxes, ensuring that file modifications, package installations, and system commands remain quarantined from host development machines.

Licensing tiers, commercial usage, and Mini-SWE-agent pricing

Mini-SWE-agent is fully open source under the MIT license with zero commercial restrictions:

*Pricing and plan details verified as of August 2026.

Step-by-step installation and local terminal onboarding guide

1. Install via pip: Install the official package using Python package manager: pip install mini-swe-agent.
2. Configure environment variables: Set your target model API key (e.g. export ANTHROPIC_API_KEY=your_key or export OPENAI_API_KEY=your_key).
3. Run interactive session: Launch the agent in your project directory by typing mini in your terminal.
4. Execute batch evaluations: Run mini-swe-agent-swebench across benchmark datasets or launch the trajectory inspector to analyze step-by-step agent execution logs.

Editorial verdict

• Best for: AI researchers, benchmark evaluators, and developers looking for a fast, radically simple, and highly transparent coding agent that executes tasks through standard bash.
• Not recommended for: Non-technical users seeking a heavy visual GUI with drag-and-drop workflow builders.
• Learning curve: Very low. The 100-line architecture can be read and understood in minutes; CLI execution is immediate.
• Value threshold: Unbeatable. Clean, open-source MIT code with zero platform overhead or subscription locks.
• Bottom line: Mini-SWE-agent is a masterclass in agentic engineering, proving that minimalist architectures deliver faster startup times, higher benchmark stability, and uncompromised coding power.