In practice
In practice it looks like this: you write down what should be true, hand it to Claude Code, Cursor, Codex or Grok, let it run with tests, and then read the diff and the evidence before anything merges. The agent does most of the typing. You own the architecture, the quality and the decision to ship.
Search results use the term in a few overlapping ways, and most are blogs and vendor pages. The common meaning matches Simon Willison's: "developing software with the assistance of coding agents." Addy Osmani puts the split this way: "AI does the implementation, human owns the architecture, quality, and correctness." IBM credits Andrej Karpathy with the term. A broader use also exists, meaning designing and running systems of AI agents, and an academic paper proposes "Structured Agentic Software Engineering" as a research agenda.
How it differs from the neighbors. Vibe coding means prompting and accepting code you do not review. Willison wants that term kept for "unreviewed, prototype-quality" code. Agentic coding, the term for the agent doing multi-step work in your codebase, describes the tool and the loop. Agentic engineering describes the discipline around it: specs, checks, review and ownership. See [agentic coding](/glossary/agentic-coding).
How to start
- Write each job as a spec: why, the evidence, what is out of scope and Done When lines an agent can check.
- Keep your standing rules in one file every agent reads, such as AGENTS.md.
- Give every job a check the agent can run: a test, a build or a screenshot.
- Read the diff and the evidence yourself before you merge.
- Choose what to build from customer evidence, not from what the agent suggests.
Common mistakes
- Using the word for vibe coding with a nicer name. If nobody reviews the output, it is not engineering.
- Writing good specs and no checks. A spec the agent cannot verify is a wish.
- Letting each agent find its own work. Several agents with no shared plan overlap or drift.
Why it matters for product builders
The word moves the focus from how much the agent can write to how you know it is right. A spec an agent can check, a test that runs and a person who reads the evidence are the work.
It also moves a question to the front that no tool answers on its own: what should the agents build? Faster agents build the wrong thing faster when the direction is gut feel.
What changes in AI-first products
This is an AI-first practice by definition, so there is no older version to compare. Much of it is old engineering habit made necessary: written requirements, tests, code review, a clear definition of done.
Founders running two or more agents add one more part: direction. Annsa holds ranked priorities, the evidence behind each and specs with Done When lines, and Claude Code, Cursor, Codex and Grok read them over MCP. They claim work and record evidence against each line. You review it and decide.
See how Annsa does it: Direct your agents →