Harness Engineering Primer
Hook
Picture an agent given real permission to open pull requests against a production codebase. Left completely alone, that's terrifying — it could rewrite half the repo in one commit, invent an API that doesn't exist, or quietly skip the test suite because nobody told it not to. A harness is the scaffolding built around that agent so its freedom to act stays useful instead of becoming a liability. This course assumes you've already built one and know its five moving parts by name.
Explanation
Harness Engineering organizes the scaffolding around an agent into five jobs, and each one answers a different question about what the agent is allowed to do.
Constrain narrows the space of actions before the agent ever takes a turn — a diff size cap, a list of files it's not allowed to touch, a rule that it can propose a merge but never execute one itself. Constraint is prevention: it removes a category of mistake by making it unreachable, rather than catching it afterward.
Inform hands the agent the context it needs to act well — the team's style guide, the ticket it's working from, the result of the last test run. An agent that constrains without informing just produces safe, useless output; informing is what makes the constrained space worth acting inside.
Verify checks the agent's output against something outside the agent's own opinion of itself — a test suite passing, a schema validating, a lint rule holding. This is the step that stops "I'm confident this is right" from being treated as evidence that it actually is.
Correct takes a verification failure and, where the fix is narrow and well understood, applies it automatically — reformatting a file, retrying a flaky call, filling in a missing field — so a human isn't paged for something a script could have handled.
Escalate is the honest fallback: when a failure doesn't fit a known correction, the harness stops pretending it can handle everything and hands the decision to a person, with enough context attached that the handoff doesn't cost them another hour of digging.
Check yourself
An agent's tests fail after it opens a pull request. The harness reformats a misindented line automatically, then still can't get the suite green, and pings a human with the failing test output attached. Which job did the reformat belong to, and which job did the ping belong to?
Reveal the answer
The reformat is a correction — a narrow, understood fix applied without asking. The ping is an escalation — the harness recognizing the remaining failure doesn't match anything it knows how to fix on its own, and handing it off with the context a person needs to act on it.
Hold on to that last job in particular: the shared graph this course teaches you to build is exactly the kind of outside evidence a harness's verify stage reaches for when "trust the agent's own report" isn't good enough anymore.