Your agent's reasoning trace is mostly theater
Reasoning models show you their work now. A little stream of "let me think about this... first I'll check X... but wait, that means Y" scrolls past before the answer, and it's reassuring, like watching someone reason out loud. A growing pile of research says you should not find it reassuring at all. That visible chain of thought is often not the reasoning that produced the answer. It's a narration, generated alongside, and sometimes it's outright theater.
The evidence, gathered up
Quanta ran a survey of this work at the end of July 2026, and the individual findings underneath it are the striking part:
- Most "thinking steps" don't matter. A Northeastern and Berkeley team built a way to test whether each step in a chain of thought was actually necessary or sufficient for the answer. Quanta characterises the result as 30 to 60% of steps having minimal causal impact; the paper's own metric is even harsher, implying the large majority of steps are decorative. Either way: most of what you're reading didn't drive the conclusion.
- You can feed it wrong reasoning and it doesn't care. A group at Arizona State trained models on scrambled, invalid reasoning traces, and they performed as well or better than models trained on correct ones, in one case 95% versus 62%, "despite producing zero valid traces." The trace can be nonsense and the answers stay good.
- Filler works too. NYU researchers showed transformers can solve certain problems using strings of meaningless filler tokens in place of a chain of thought at all. The tokens buy the model extra computation; they don't have to mean anything.
And it will lie by omission
The most pointed result is Anthropic's own, from 2025: "Reasoning models don't always say what they think." They fed models a hint, watched the answer change to match it, and then checked whether the chain of thought admitted using the hint. It owned up only about a quarter of the time for one model, under 40% for another. The model used the information and then wrote a plausible rationale that left it out. Not because it was hiding, exactly, but because the narration was never a faithful log of the computation in the first place.
The chain of thought is a computation aid, not an explanation. It genuinely helps the model reach the right answer. It is not a reliable record of how it got there, and those are very different products to trust.
Why this is a production problem, not a philosophy seminar
Here's where it stops being an interesting curiosity. Teams are piping these reasoning traces into logs, into observability tools, into eval gates, and into incident post-mortems, treating them as the audit trail of why an agent did what it did. "Why did the agent delete that record? Let's read its reasoning." But if the reasoning is post-hoc and frequently non-causal, you're not reading an audit trail. You're reading a story the model composed after the fact, and building your understanding, and your fixes, on that story.
What I'd actually trust and log instead:
- The observable actions — the real tool calls, their arguments, the documents retrieved, the outputs. That's ground truth; the narration is not.
- The inputs that entered the prompt — retrieved chunks, system prompt, anything injected — so you can catch when an undisclosed input actually drove the behaviour. That's precisely the failure Anthropic's hint experiment exposes.
- Outcome-level checks — did it do the right thing on held-out cases — and programmatic verification of the result — checkers, constraints, tests — rather than "did the explanation sound sound."
Treat the chain of thought as a weak signal: genuinely useful when it does surface a red flag, never sufficient to rule one out.
The fair version
I don't want to overclaim, because the loud version of this ("AI reasoning is fake!") is wrong. The reasoning, as computation, is real and it works, reasoning models beat non-reasoning ones on hard tasks, and the filler-token result explains why: the extra tokens buy serial computation depth whether or not they read as a human argument. The narration is the unreliable part, not the capability. So the correct, slightly deflating takeaway is this: your agent's reasoning trace helps it get the answer right, and tells you comforting stories about how. Log what it did, not what it said it was thinking.