SWE-agent (from Princeton NLP) is a coding agent designed to fix real GitHub issues. It runs in a containerized agent-computer interface with a custom shell tailored for software engineering tasks. Strong benchmark performance on SWE-bench. Sandboxed by design.
Autonomy
Semi-autonomous
Sandbox-aware
Yes
Network access
Restricted
Filesystem access
Restricted write
Permissions declared
Yes
Pattern
Single agent
MCP servers used
Models
gpt-4oclaude-3-5-sonnet