Reasoning Alone Is Brittle

The Plan That Breaks

A language model writes a confident plan: search, filter, open the product, compare prices.

Then one click returns a page that does not match the plan.

The model keeps going anyway. Nothing in its head told it the world had changed.

Notice that the plan on the left keeps its original order while the page on the right has already changed, so nothing corrects the mismatch.

This is the gap between thinking about a task and doing one.

What A Reasoning Trace Does

A reasoning trace is the model's intermediate thinking written out before or between actions.

It tracks the goal, the constraints, the subgoals, and the candidate next step.

Chain-of-thought prompting is one familiar version of this.

Traces are useful because they expose assumptions. Someone can read them and see where the plan came from.

Fluent Is Not Correct

Here is the trap. A trace can be detailed, well organized, and completely wrong.

Fluency is not evidence. The model may be planning, tracking what it has learned, or quietly rationalizing a choice it already made.

Reading traces well means asking which of those three is happening.

A rationalizing trace looks the most convincing and helps the least.

Why Plans Go Stale

Interactive environments disagree with plans constantly.

A web page changes layout. A search result comes back irrelevant. A command fails. A required fact is simply missing.

None of these appear in the model's head. They only appear in what the environment returns.

So a longer, more careful plan written up front does not remove the need to look.

Reason, Act, Observe, Revise

ReAct's proposal is simple. Interleave reasoning traces with task-specific actions instead of studying them separately.

The trace induces a plan. An action tests it. The observation comes back. The trace updates.

That last arrow matters most. Reasoning traces help track and update plans and handle exceptions.

Without the return trip, an agent is just narrating.

Where Each Half Fails Alone

Reasoning without acting is brittle because the world can disagree with the plan.

Acting without reasoning is brittle because actions need goal-directed selection.

There are honest boundaries here. A pure essay answer can reason with no actions at all.

A one-click macro can act with no reasoning at all. Neither is an agent operating in a changing environment.

Diagnosing A Failure

When a computer-use agent fails, the useful question is which part broke.

Was reasoning missing, so the action had no goal behind it? Was observation missing, so the agent never saw the result?

Or was revision missing, so it saw the mismatch and continued anyway?

The shopping agent that opens the wrong product and keeps comparing prices failed at revision, not at planning.

Is a fluent reasoning trace automatically correct?

• Yes
• No

Answer: No

No. Fluency and detail show how convincing the text sounds, not whether its assumptions or conclusions match reality.

In the ReAct cycle, the agent first _____ about a goal, then _____ in the environment, _____ the returned result, and _____ the plan.

Answer: In the ReAct cycle, the agent first reasons about a goal, then acts in the environment, observes the returned result, and revises the plan.

ReAct connects internal planning to external feedback: reason, act, observe, and revise. Revision lets the agent respond when the environment differs from its expectations.

A shopping agent writes, “The filter will show the cheapest shoes,” before clicking anything. What is this reasoning trace mainly doing?

• Planning a next step from an assumption
• Tracking a result already observed
• Rationalizing a choice after it was made
• Revising the plan after an error

Answer: Planning a next step from an assumption

The trace is planning because it proposes a next step and predicts an outcome before receiving environmental feedback. The prediction may still be wrong.