What is agentic reasoning?
Reaching a conclusion over multiple steps, where those steps are able to alter the available facts. Deliberation itself is nothing new; language models have produced intermediate steps for years. What separates this is that the deliberating can reach outside itself and retrieve what was missing.
The word is used for two different mechanisms, and keeping them apart resolves most of the confusion around the term.
| Mechanism | What happens | What it cannot do |
|---|---|---|
| Model-internal reasoning | The model generates intermediate steps before answering, spending more computation on a fixed input | Acquire anything it was not given. Thinking longer about incomplete information yields a better-argued answer from incomplete information |
| Agentic reasoning | The system decides an action, takes it, observes the result, and reasons again with what came back | Escape a wrong framing. It can gather evidence and still gather the wrong evidence, repeatedly and convincingly |
Most production systems use both, which is why they get conflated. A capable reasoning model inside a loop that never retrieves anything is still answering from what it already had, and that configuration is common in deployments described as agentic.
How is Agentic Reasoning different from a reasoning model?
By what it can change. A reasoning model improves the quality of thinking applied to an input it cannot alter. An agent treats the input as something it can go and extend, which makes the limiting factor what it is able to reach rather than how well it thinks.
That difference has three practical consequences.
- The failure moves from the model to the connections. An agent that cannot reach the right system produces a well-reasoned answer from the wrong material. The quality ceiling is set by access, which is the argument at enterprise connectors.
- Cost becomes unpredictable. A reasoning model has a bounded cost per call. An agent decides how many calls to make, so the same question can cost very differently on two occasions depending on what it finds.
- It can be wrong for longer. A model answers once. An agent can pursue a mistaken line across many steps, each individually sensible, and arrive somewhere confidently incorrect having spent considerably more to get there.
The upgrade that actually matters is unglamorous. Reasoning improves most when the agent can check something rather than when it thinks harder, so the highest-return intervention is usually connecting it to a source of truth rather than selecting a more capable model.
What are the main reasoning patterns?
Four, and they differ on when the thinking happens rather than on how much of it there is. That timing determines whether the work can be reviewed before anything happens, which is the property that matters in a regulated process.
Each with the trade it makes.
- 01
Reason, act, observeThe agent alternates between thinking and doing, one step at a time, deciding the next move from what the last one returned. Trade: highly adaptive, and nothing is reviewable in advance because the path does not exist until it is walked.
- 02
Plan first, then executeA full plan is produced before anything runs. Trade: the plan can be approved, logged and compared against what actually happened, at the cost of adapting poorly when reality differs from the plan.
- 03
Reflection and self-critiqueThe agent reviews its own output and revises. Trade: catches format and consistency errors reliably, and is weak on factual errors, because the thing checking shares the beliefs of the thing being checked.
- 04
Search over alternativesSeveral lines are explored and the best is selected. Trade: genuinely better on problems with a verifiable answer, and expensive, with the selection step as the weak point when correctness cannot be checked mechanically.
The second is the one to reach for when anything consequential happens, for a reason that has nothing to do with quality. A plan produced before execution is the only artefact a human can approve, and approval before the fact is a different control from review after it. That distinction drives the argument at agentic workflow.
When does more reasoning make things worse?
On simple tasks. Longer reasoning is often sold as a dial that only improves answers, but it adds cost and delay. It can also hurt accuracy, because the AI may talk itself out of a correct first answer. Save deeper reasoning for hard, multi-step problems.
Four situations where additional steps hurt.
- The task is simple. A lookup with one correct answer gains nothing from deliberation and can lose by it, as additional steps introduce opportunities to go wrong that a direct answer did not have.
- The premise is wrong. Reasoning compounds from its starting point, so a mistaken framing produces a longer and more persuasive journey to the wrong place. Length increases confidence without increasing correctness.
- Nothing can verify the conclusion. Search and reflection improve results where an answer can be checked. Without a check, extra passes select for plausibility rather than accuracy.
- The context is already crowded. Each reasoning step adds to what the model must attend to, and beyond a point the earlier material gets crowded out, which is covered at context engineering.
The failure this produces has no good name and deserves one. An agent reasons from a wrong premise, every individual step is locally valid, the conclusion is wrong, and the trace reads beautifully because the logic genuinely is sound. Nothing in the record looks like an error, so review passes it. More reasoning makes this worse rather than better, since each additional step is more sound logic built on the same faulty base.
Why does a good trace not mean good reasoning?
The account and the computation are separate artefacts, and only one of them is a recording. What appears as a chain of thought is text the model wrote, and text written about a decision need not match the mechanism that actually drove it.
Three implications for anybody relying on a trace as evidence.
- Fluency is not fidelity. A well-written chain of thought demonstrates that the model writes well about reasoning, which is a different capability from reasoning well.
- What the agent did is reliable; why it says it did so is not. Tool calls, retrievals and actions are observed events. The narrative connecting them is produced afterwards and should be read as commentary.
- Evaluating the trace rewards articulacy. Scoring reasoning quality by reading the explanation selects for systems that explain persuasively, which is not the property you were trying to measure.
This page deliberately stops here, because the full argument belongs elsewhere. Neither the execution trace nor the stated reasoning is an explanation, and what that means for audit, regulation and the right to an explanation is set out at agent explainability.
How do you evaluate Agentic Reasoning?
On outcomes and on the observable actions, not on the quality of the account. The instinct is to read the reasoning and judge whether it sounds sensible, which measures how convincing the system is rather than how often it is right.
Five measures, and the first two carry most of the signal.
-
Outcome correctness on cases with a known answer
Unglamorous, and the only measure that is not a proxy. Build the set from real past cases where somebody established the right answer, including the awkward ones.
-
What it retrieved, not what it concluded
For a wrong answer, check whether the agent reached the right source. Wrong source is an access problem; right source and wrong conclusion is a reasoning problem. Those need entirely different fixes.
-
Steps taken against steps needed
An agent using fourteen steps for a three-step task is reasoning badly even when the answer is correct, and it is doing so at roughly five times the cost.
-
Behaviour when the answer is not available
Give it a question your sources cannot answer. Saying so is a correct outcome; constructing a confident response from adjacent material is the failure that matters most in production.
-
Consistency across reruns
The same question several times. Divergent reasoning paths reaching the same answer is healthy. Divergent answers mean the reasoning is not doing the work you think it is.
One note on using a model to judge reasoning, since it is the common shortcut. A model grading another model's reasoning tends to reward the style it would have produced itself, which inflates scores for articulate wrong answers. It is useful for throughput and needs calibrating against human-labelled cases, which is the argument at agent evaluation.