Agentic Is a Loop That Closes, Not a Label on a Chatbot
Pull-quote: “If a system cannot notice that its own action failed, it is not an agent. It is a script with a good vocabulary.”
The word stopped carrying information
The word agentic now appears on products that plan multi-day investigations and on products that call one API before replying. When one label covers both, the label carries no information, and evaluation decisions get made on adjectives. We use a narrow definition when we build and a short test when we evaluate. An agent is a system that pursues a goal by iterating a loop, perceive, reason, act, verify, until a stop condition it can check for itself. Every part of that sentence is load-bearing, and most systems sold as agents fail at the fourth stage.
The loop, stage by stage
PERCEIVE ──► REASON ──► ACT ──► VERIFY ──► stop check ──► done
read the revise typed check │
world the plan tool outcome │ not done, or
fresh call at the │ verify failed
▲ source │
│ │
└───────────────────────────────────────────┘
loop again, with new evidence| Stage | Question it answers | Minimum machinery | Common counterfeit |
|---|---|---|---|
| Perceive | What is true right now? | Fresh reads of world state, provenance on inputs | A static prompt assembled once at the start |
| Reason | What should change next? | A revisable plan tied to goal and current state | One chain of thought generated at step zero |
| Act | Make the change | Typed, permissioned tool calls with declared side effects | Prose that claims an action happened |
| Verify | Did it work? | An independent check against the source system | Asking the model whether it is confident |
Two stages do the discriminating. Perceive means reading the world again, not reasoning over a snapshot assembled when the run began. Verify means checking the outcome at the source system, independently of what the transcript claims happened. A model narrating “the record has been updated” is not verification. A fresh read of the record is.
Three questions that expose the difference
We use this test in evaluations, and we accept the same test in return.
- Show the step where the system observes the result of its own action. The world state after the action, not the tool’s return code.
- Show one run where verification failed and the plan changed because of it. If no such run exists, the loop has never closed.
- Show what stops it. A budget, a done-check, an escalation rule. If the honest answer is that the user closes the tab, it is a chatbot.
A system can fail this test and still be useful. Retrieval assistants and single-shot tool callers earn their keep every day. The failure is not the architecture. The failure is selling the architecture as something it is not.
The definition also pays off in engineering, not just procurement. Once you commit to the loop, you evaluate runs instead of responses: did the plan converge, how many verification failures occurred, what stopped the run. Budgets and stop conditions become testable requirements rather than aspirations, and an incident review can point at the stage that failed instead of debating whether the model was smart enough.
What this looks like in practice
Investigation and casework agents show the loop at full length: after each action, the case state is re-read before the plan advances, so verification happens against the data rather than the narrative, and a failed check sends the plan backward instead of forward. The discipline also runs in the other direction. A public-facing assistant is often best built as a short-loop agent with narrow tools and tight stop conditions, because a system that acts in front of strangers earns trust by being scoped and honest about it.
Closing
Agentic is not a model tier and not a pricing page. It is an engineering commitment: read the world fresh, revise the plan, act through typed tools, verify at the source, and stop on your own terms. Systems that do this exist and deserve the label. For everything else there is an older, perfectly respectable word: software.