What is agentic AI? The term, minus the hype

Agentic AI describes systems that pursue goals by taking their own sequence of actions, rather than producing a single response and stopping. It’s an adjective about degree — how much of the decision-making has moved from the person to the system — which is why it gets attached to almost everything. Here’s a scale that makes it useful.

Five levels of agency

Most arguments about whether something is “really agentic” dissolve once you place it on a scale instead of asking a yes/no question.

Level 0 — Generation. The model produces text or an image. You do everything with it. Autocomplete, drafting, summarising a document you pasted in.

Level 1 — Retrieval. It fetches context before answering. RAG lives here. The system reads more than you gave it, but still only answers.

Level 2 — Tool use. It can call a function you exposed. One call, inside a turn you initiated, chosen from a short list. Most “AI features” shipped in the last two years are here.

Level 3 — Multi-step execution. It chains calls, reads what came back, and adapts. This is where the word agent starts being honest: the sequence wasn’t written by a human.

Level 4 — Delegated goals. It runs against an objective over time, on a trigger or a schedule, deciding what’s worth doing and escalating when it isn’t sure. This is where the practical value is, and where the engineering gets hard.

Level 3 is achievable and reliable today. Level 4 works when the domain is bounded and the escalation rules are good. Anything claiming full autonomy across arbitrary tasks is overselling.

Why it became the word of the year

Two capabilities crossed a threshold at roughly the same time.

Models got materially better at structured output — reliably emitting a well-formed tool call rather than prose describing one. That made the act step dependable, and an unreliable act step means no loop at all.

And long-context reasoning improved enough that a model could hold a plan, the results so far, and the remaining goal simultaneously. Before that, multi-step attempts drifted: by step four the model had lost track of step one.

Neither is a new idea. Agent architectures were being written years earlier. What changed is that the substrate finally supported them.

The part vendors skip

Agentic systems fail differently from chatbots, and the failure modes are what determine whether you can deploy one.

Compounding errors. A 95%-reliable step is fine. Ten of them in sequence is 60%. Long chains need checkpoints, not just better prompts.

Confident wrong actions. A chatbot that’s wrong produces a bad paragraph you can ignore. An agent that’s wrong sends the email. The consequence of an error scales with the permissions you granted.

Unbounded cost. A loop that retries can burn budget on something that was never going to work. Caps and circuit breakers are load-bearing, not polish.

Unauditable runs. If you can’t replay what it did and why, you can’t debug it, and you certainly can’t answer a compliance question about it.

None of these are reasons to avoid agentic systems. They’re the actual engineering work, and a vendor who doesn’t discuss them hasn’t hit them yet.

What it means for buyers

Ignore the label and ask what level you’re being sold. A product at level 2 described as “agentic” isn’t lying, exactly — it’s just answering a much easier question than the one you’re trying to solve.

Then ask the three questions that separate levels 3–4 from the rest:

• Who chooses the steps? If a human drew the graph, it’s automation.
• What happens on failure? Adapting is level 3. Stopping or pretending is level 2 with ambition.
• What can it do unsupervised? The answer should be a policy — reads yes, irreversible writes no — not “everything” or “nothing”.

Where we sit

Velaris is a level 3–4 system with a deliberate ceiling: it plans and chains its own steps across a thousand tools, runs on triggers and schedules, and stops for a human on anything outbound or irreversible.

That ceiling is a product decision, not a limitation we’re working around. An agent you’d actually grant access to your inbox is more useful than a more autonomous one you wouldn’t.

Start with what an AI agent is, or see one running.