What "agentic" actually means to us

"Agentic" has eaten itself. Every product page uses it. Most mean something different. Nobody asks anymore, so we just nod.

Our internal definition

A system is agentic when three things are true:

1. It picks its next action. Not a hard-coded pipeline. A step depends on what just happened.
2. It has tools. Reading a file, calling an API, opening a browser — the system acts on the world, not just its own tokens.
3. It continues across multiple turns without being poked. A single-shot tool call is not an agent. An agent is a loop.

By this definition, most "AI features" are not agentic, and that's fine. A well-written assistant that opens a file on your behalf is not agentic — it's a command. We save the word for the loops, because the loops behave differently: they compound mistakes, they rack up cost, they need different evaluation.

Why this matters for your team

If you don't pick a definition and stick to it, every conversation about "building an agent" means something different to each person at the table. You end up budgeting for a demo and shipping a pipeline. Pick a definition. Let it be stricter than you're comfortable with. You'll make better decisions.