A reasoning model is a language model trained or configured to work through a problem in intermediate steps before giving its answer, spending more computation per request in exchange for better results on multi-part tasks. It tends to do better at calculations, planning and applying layered rules. It is slower and costlier per call, and it is no better at judging whether it is allowed to act.
When does an agent benefit from a reasoning model?
On tasks where getting one step wrong spoils the result. Working out which of several fee rules applies to an account with a complicated history. Planning a sequence of steps for an orchestrator. Reconciling two sets of figures that differ for several reasons at once. Reading a long contract to find which clause governs a situation.
For simple steps, such as classifying an email or pulling a date from a form, a reasoning model adds cost and delay without much gain. Most agents use a mix.
Why is better reasoning not the same as safer action?
A reasoning model is better at reaching the right answer to the question it was asked. It is not better at knowing whether the action it proposes is within its authority, whether a fact it relied on came from a forged email, or whether the customer's limit has changed since the agent was set up. Those are questions about bounds and evidence, not intelligence.
It also brings a subtler risk: persuasive working. A reasoning model can produce a long, coherent explanation for a wrong conclusion, and the length makes it more convincing to a reviewer. Its written steps are generated text, like chain-of-thought reasoning generally, not a guaranteed account of why it acted.
How is a reasoning model used in practice?
Teams route the hard steps to it and keep the rest on faster models. The reasoning model's output, like any model's, is a proposal. Actions it proposes go through the same checks as any other agent action. Its working can be kept with the case for review, treated as context rather than evidence.
For example, an asset manager's operations agent reconciles a fund's positions against the custodian's report. Most lines match and are handled by a smaller model. The handful that do not match go to a reasoning model, which works through corporate actions, settlement timing and fees to explain each break. It proposes adjusting entries; a person in fund accounting approves them.
How is a reasoning model different from a foundation model?
A foundation model is any large model trained on broad data for many tasks. A reasoning model is a foundation model built or set up to spend extra effort working through steps. All reasoning models are foundation models; most foundation models used day to day answer directly. The difference is in how the model approaches a request, not what it is.
Can a reasoning model supervise other agents?
It can review their work and catch some errors, which is useful. It should not be the final authority on whether actions proceed, for the reason set out in why the model cannot be its own supervisor: a model checking a model shares the same blind spots, including susceptibility to injected instructions. Because reasoning models change and improve quickly, controls should not depend on any one of them; see whether an agentic operating system can work with any model and switching model provider without losing control.
Questions people ask
Are reasoning models less likely to hallucinate?
On some tasks, yes, because they check intermediate steps. They can still invent facts, and their longer outputs give more room to do so. Grounding and checks against source records still apply.
Should every agent use a reasoning model?
No. Use one for steps where layered logic matters. The extra cost and delay are wasted on simple steps. The governed agentic operating system guide explains why the controls stay the same either way.