Agentic AI GlossaryReading guide

Curated reading path

Your first coding agent

Learn what controls the work, how tools act, and what evidence to ask for before accepting a coding change.

Start with a small, verifiable change

For readers who have used chat but have not yet delegated a repository change. Read these entries in order; each link opens the term within the continuous glossary. Use Back to return to this sequence.

By the end, you should be able to distinguish the model from the system executing its requests, set a bounded task, and ask for evidence of completion. Keep one example in mind: changing an empty search message while preserving keyboard navigation.

Reading order

  1. Model

    Identify the trained component before attributing filesystem access or persistence to it.

  2. Harness

    Locate the software that assembles context and executes authorized actions.

  3. Agent

    See how a model-directed loop can inspect results and choose another step.

  4. Tool

    Understand the boundary between generating a request and changing the environment.

  5. Permission mode

    Separate the ability to request an operation from authority to execute it.

  6. Turn

    Follow multiple tool and model steps within one user-visible exchange.

  7. Automated check

    Ask for checks that cover the requested behavior, not only a green status.

  8. Human review

    Inspect the actual artifact and unresolved risks before accepting the change.

Put the sequence to work

In a disposable repository copy, describe the expected behavior, identify the allowed files, and give the relevant test command. Inspect the diff and ask which requirement each check covers. You are ready to move on when you can explain why an applied patch alone does not establish that keyboard navigation still works.