Without verification you only have the agent’s word that the work is done.
Verification turns that word into a fact. Sometimes it is a test going green. Sometimes it is a rule you wrote down beforehand, like 4 out of 5 on our UX scale.
That is where quality comes from. Once the rule is written down, the agent verifies its own work against it, and keeps looping until it gets there.
Three things carry the method, and the update works each one through in full.
Define what done looks like, in a simple manner like in one sentence. Write it so the finished work can be judged against it rather than argued about.
Define how it gets verified. As an example, a command that returns a fact: a linter reads the code and passes or fails, or a step where an LLM looks at the finished page and returns a judgement, scored 1 to 5, where 4 or above passes.
Say to keep going until both pass, then stop and say what changed. Nothing else is set up: no plugin, no script, no settings file. The instruction is the loop.
One small web page with four problems planted in it, two verifications, and the loop that runs them. You copy plain text files into your own project. There is no installer and nothing to configure.
Takes a change request, makes the change, runs both verifications, and repairs what failed.
Opens the finished page in a browser and scores it against the rules you wrote down. MET or NOT MET, with what it saw quoted.
Runs the linter your project already has — the tool that reads the code and flags what is wrong with it. PASS or FAIL, with the file and the line.
Where you watch it happen. The linter catches two of the planted problems. Only the finished page shows the other two.
The whole example, plus four exercises that put each piece of it into a codebase you own. The fundamentals of agentic work are taught in our classroom course, which is a separate thing.