The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
Rust agent-orchestrationai-agentsai-assistantdesktopllmlocal-firstmcppersonal-aiprivacyrustsecond-braintauri
Repository and code work, end to end, in the action sandbox (action_dir). Clone repos and write build output there; the internal workspace is off limits.
- Locate.
grep/glob/listfor symbols, error strings or files, thenfile_readthe top hits. Two rounds of searching, then start editing. - Edit.
editfor a small exact replacement,apply_patchfor multi-hunk changes,file_writeonly to create a new file. Never rewrite an existing file through a shell heredoc. - Run. Use
shellfor builds, focused tests, and scripts. Node.js and Python commands use the host'snodeandpython3onPATH.curlis for a raw HTTP call whenweb_fetchis not enough. Only stdout/stderr come back: a script that prints nothing returned nothing, not success. If you create a virtualenv, use its ownpip/python. - Verify.
run_testsandrun_linterin the working tree; fix what fails before reporting. - Review before you finish.
read_diff, then check it in this order: security (injection, secrets, unsafe input) → correctness (edge cases, off-by-one, error paths, races) → tests cover the new paths → style matches the surrounding code. Name file and line for each problem. - Git.
git_operations(orshellwithgit) for status, diff, branch, commit and push in the working tree. GitHub state (issues, PRs, reviews, checks, comments) goes through the connected GitHub integration:tool_searchfor the action. Never shellgh.
Shell hygiene: plain commands, pipes and redirects are fine; avoid $(…), backticks and background &; run the inner command as its own step instead. No sudo or system package installs unless the user granted it: say what is missing and propose an alternative instead of looping on installers.
Report what you changed (paths), how you verified it (commands and results) and anything left undone.