What agentic actually means

Ask ten engineering leaders what "agentic" means and eight will describe something that isn't new. A chat window that writes better code than it did last year.

That's not nothing. It's also not the thing driving the current wave of tools, budgets, and anxious board questions.

The distinguishing feature is a loop.

An agent reads a goal, decides what to do, takes an action (edits a file, runs a command, calls an API), looks at what happened, and decides what to do next. No person approving each step. Autocomplete finishes your sentence. Copilot suggests your next line. An agent opens the file, makes the change, runs the tests, reads the failure, and tries again, four or five times, before it ever shows you a diff.

That loop is the whole distinction.

Worth being precise about, because a lot of procurement decisions right now are getting made on vibes rather than on this one mechanical difference.

Three things have to be true

For "agentic" to mean anything beyond a marketing page, three things need to hold at once.

It needs a goal, not just a prompt. Something like "make this test pass" or "migrate this file to the new API." Something it can check its own progress against.

It needs tools. A shell, a file system, an API. A way to act in the world rather than only producing text that a person then has to act on.

And it needs to run more than one step without you. Choosing the next action itself, based on what the last one returned, until it finishes or gets stuck.

Remove the loop and you have a smarter autocomplete. Remove the tools and you have an unusually capable rubber duck. Remove the goal-directedness and you have a chatbot that's good at code.

All three together, running unattended for minutes at a stretch, is the actual novelty. And it's newer than most people assume. Reliable, multi-step tool use in coding contexts is mostly a 2024 and 2025 development, not a 2022 one.

Why the distinction is worth your time

You could treat this as pedantry. A distinction that matters to model vendors and nobody else.

It doesn't hold up that way, because the loop is exactly where both the value and the risk live.

The value first. A tool that runs ten minutes unattended, checking its own work against tests, doing the tedious verification a person would otherwise do by hand, hands back something closer to a finished draft than a suggestion.

That's a different kind of help than typing faster.

Then the risk, which is the same mechanism pointed the other way.

Ten unattended minutes is ten minutes where a wrong assumption compounds before anyone notices. A misread of what a function does gets baked into forty files instead of one. A destructive command runs because it was the fastest path to the stated goal.

What the agent is allowed to touch while nobody's watching turns out to be the entire subject. It deserves its own lesson, not a paragraph here.

The 2023 demos of "AI agents" mostly weren't. Chained prompts with a lot of narration, falling over after two or three steps on anything real.

Writing off the category on that basis was a reasonable read of 2023 and a wrong one by the middle of 2024, once tool calling got reliable and models got good enough at noticing their own mistakes to correct course.

Which is why the useful question is never whether something is AI. It's whether the thing loops without you, and on what.

The question to ask about any tool someone wants to buy

Skip the marketing page.

Ask three things.

What's the goal it's working against, and can it tell when it's met. A tool with no way to check its own output isn't agentic. It's just fast.

What can it actually touch. Read-only tools that suggest changes sit in a different risk category entirely from tools that can commit, deploy, or call a payment API.

How many steps before a person looks at anything. One step per check is a very different product, and a very different risk profile, than fifty.

What this means for evaluating tools this year

Every vendor now says "agentic."

Most of what ships under that label is still closer to autocomplete with a longer memory. A smaller number genuinely run the loop. Plan, act, observe, decide, repeat, with real tools and real consequences attached.

That isn't academic once you're deciding what an eight-person startup should adopt against what a four-hundred-person org should let near production.

Ask what a tool is allowed to touch before you ask how good the model is. The model is the part vendors compete on. The loop, and what you let it run against, is the part you're responsible for.