SWE-agent — Automated Bug Fixing with AI Agents

How It Works

1. Issue Analysis

SWE-agent reads the GitHub issue and understands:

• What is broken
• Expected vs actual behavior
• Relevant error messages and stack traces

2. Codebase Navigation

Using its Agent-Computer Interface (ACI):

# The agent navigates like a developer
find_file "auth_handler.py"
open_file "src/auth/auth_handler.py"
search_file "validate_token"
goto_line 145

3. Bug Identification

The agent reads relevant code, traces the logic, and pinpoints the issue:

Analyzing: src/auth/auth_handler.py:145
Issue: Token expiry check uses UTC timestamp but comparison uses local time
Fix: Convert both to UTC before comparison

4. Patch Generation

Writes and applies the fix:

# Before
if token.expiry < datetime.now():

# After (SWE-agent fix)
if token.expiry < datetime.now(timezone.utc):

5. Validation

Runs existing tests to verify the fix:

Running test suite...
test_auth_valid_token: PASS
test_auth_expired_token: PASS (was FAIL)
test_auth_refresh: PASS
All tests passing.

Agent-Computer Interface (ACI)

Custom commands designed for code navigation:

SWE-bench Results

Key Stats

• 15,000+ GitHub stars
• Princeton NLP research
• 12.5% SWE-bench resolve rate
• Custom Agent-Computer Interface
• Supports Claude and GPT-4

FAQ

Q: What is SWE-agent? A: SWE-agent is an AI system from Princeton NLP that autonomously fixes real-world GitHub issues by navigating codebases, writing patches, and validating fixes.

Q: Is SWE-agent free? A: Yes, open-source under MIT license. You need LLM API keys.

Q: Can SWE-agent fix any bug? A: It resolves about 12.5% of real-world software engineering tasks. It works best on well-specified bugs with clear reproduction steps.