Harness engineering for agents-cli
agents-cli is the meta harness engineering system for agents. These pages explain what that sentence means and how to use it.
Where to start
conceptsHarness engineeringWhat a harness is, who coined it, and why agents-cli is the meta layer above it.Start here if the word 'harness' is new.conceptsThe six disciplinesMitchell Hashimoto's six-step adoption arc, restated in agents-cli commands.Read after the concept page.quickstartFirst ten minutesInstall, run claude, save a session, write your first AGENTS.md. No tour, just the commands.Already convinced. Want to type something.
Recipes
One pattern per page. Each starts with the command you would type and ends with the file you would change.
Recall and resume with sessionsEvery run is captured. Use sessions to pick up exactly where the agent left off.Browser as a feedback signalScreenshots, console errors, and DOM refs are how the agent learns the UI is wrong.Parallel work with teamsSpin up four agents on four files. Boundary contracts keep them from colliding.Secrets and deployTouch-ID-gated bundles inject keys into a single command. Nothing touches disk in plaintext.
Citations
The word harness in this sense was coined by Mitchell Hashimoto in My AI Adoption Journey, published February 5, 2026. Martin Fowler's site hosts a follow-up memo by Birgitta Böckeler. OpenAI's Codex team and the LangChain engineering team published practice reports that same month. We cite all four where it matters.