In practice
You tell the agent: add Google sign-in. It reads the auth code, makes a plan, changes several files, runs the tests, fixes the failure it caused and opens a pull request. You read the diff and decide whether to merge.
Anthropic describes Claude Code as an agentic coding environment that can read your files, run commands, make changes and work through problems while you watch, redirect or step away. Cursor, Codex and Grok work in the same way.
It isn't autocomplete, which suggests the next few characters as you type. It isn't vibe coding, which is loose prompting with little checking. In agentic coding the agent loops on its own, and the checks come from tests and review.
How to start
- Write the goal as a result, with what is out of scope.
- Ask for a plan first, and read it before any code is written.
- Give the agent a check it can run: tests, a build or a screenshot to compare.
- Keep your rules in one file, such as AGENTS.md, so every session starts from the same rules.
- Review the diff and the evidence yourself before you merge.
Common mistakes
- Skipping the check. A plausible-looking change that nothing tested is the common failure.
- Starting with a vague goal and fixing it by chatting. The context fills with failed attempts.
- Letting the agent decide what to build. It can build anything; choosing is still yours.
Why it matters for product builders
The bottleneck moves. Typing stops being the slow part, and deciding what to build, writing it down clearly and checking the result takes over.
Anthropic's guidance puts it plainly: without a check the agent can run, you become the verification loop. A test, a build or a screenshot is what lets a session run without you.
Speed cuts both ways. A vague goal now becomes a finished wrong feature in an afternoon.
What changes in AI-first products
This term exists because of agents, so there is no earlier practice to compare it with. What it changes is the work around the code: specs, rules files, and tests carry more weight than they did.
What didn't change is that the agent doesn't know your customers. It builds what it is told. Founders running two or more agents feel this most, because each one needs the same direction. Annsa connects to Claude Code, Cursor, Codex and Grok over MCP so they read the same spec.
See how Annsa does it: The spec →