A coding agent is an AI agent specialised in programming: it reads a repository, understands how it is organised, changes the necessary files, runs tests and chains together dozens of modifications in a single session. Unlike an editor's autocomplete, which suggests the next line, the agent receives a goal ('change prices to whole cents', 'add this validation') and decides what to touch, in which order and how to test it.
It usually works in a cycle: it plans, edits, executes (tests, compilation, scripts), reads the result and corrects. It can open branches, prepare commits and, if permitted, deploy.
What changes with a coding agent
- Writing is no longer the bottleneck: the slow part becomes knowing what to ask for and checking that what is delivered does what it should.
- It amplifies the person directing it in both ways: it turns knowledge into finished work, and gaps into good-looking errors, repeated across many files at once.
- Permissions matter more than ever: which repositories, which branches, which credentials, and whether or not it has access to production.
Best practices
- Ask in defined steps and in the correct order, with clear acceptance criteria.
- Minimum and explicit permissions; production only with human approval.
- Tests before and after, backup and rollback prepared before deploying.
- Check the claim 'done': the claim is not the evidence.
In NAiOS, agents acting on real systems follow the human-in-the-loop (HITL) principle: they propose, and irreversible actions are confirmed by a person.