SWE-agent — Autonomous GitHub Issue Solver

What it is

SWE-agent is a tool that enables language models like Claude Sonnet or GPT-4o to autonomously fix GitHub issues, identify security vulnerabilities, and tackle custom coding challenges. It provides a specialized agent-computer interface (ACI) that gives the LLM efficient commands for navigating codebases, editing files, and running tests.

SWE-agent is for engineering teams and researchers who want to automate bug fixing and security scanning. It has achieved state-of-the-art results on SWE-bench, the standard benchmark for evaluating automated software engineering.

How it saves time or tokens

SWE-agent's custom ACI is designed to reduce the number of tokens an LLM needs to understand and navigate a codebase. Instead of dumping entire files into the context, the ACI provides targeted commands for searching, viewing specific line ranges, and editing precisely.

For recurring bug categories (null pointer exceptions, missing error handling, type mismatches), SWE-agent can resolve issues without human intervention. This frees developers to focus on design decisions and complex architecture work.

How to use

1. Install SWE-agent:

pip install sweagent

1. Fix a GitHub issue:

sweagent run \
  --agent.model.name=claude-sonnet-4-20250514 \
  --problem_statement.github_url=https://github.com/owner/repo/issues/123

1. Run on SWE-bench for evaluation:

sweagent run \
  --problem_statement.swe_bench_id=django__django-11905

Example

Running SWE-agent with a custom problem statement:

# Create a problem description file
cat > problem.md << 'EOF'
The `parse_date` function in `utils/dates.py` raises a ValueError
when the input string contains a timezone offset like '+05:30'.
Expected behavior: parse the offset and return a timezone-aware datetime.
EOF

# Run SWE-agent
sweagent run \
  --agent.model.name=claude-sonnet-4-20250514 \
  --problem_statement.file_path=problem.md \
  --repo.path=/path/to/your/repo

SWE-agent navigates to the relevant file, understands the bug, writes a fix, and validates it by running tests.

Related on TokRepo

• Coding AI tools -- AI-powered development tools
• Security AI tools -- automated vulnerability detection

Common pitfalls

• SWE-agent works best on well-defined issues with clear reproduction steps. Vague issues like 'improve performance' produce poor results because the agent cannot verify its fix.
• The agent modifies files in the repository. Run it in a separate branch or Docker container to avoid unintended changes to your working tree.
• Token costs add up quickly. A single complex issue can consume thousands of tokens as the agent explores the codebase. Set token budgets to prevent runaway costs.