How AI Works
$definereasoning--plain-english
TLDRLetting the model work a problem before it blurts.
Some questions have one obvious next file. Some questions have a trap.
A normal model commits on the first pass: next token, then the next, until it has a paragraph. That is fine for "rename the button" and rough for "why does Better Auth work locally and fail on the preview deploy?" The first path looks fluent. The first path is often the wrong fork.
Reasoning is paying for a private (or visible) work-up before the line you asked for. The model spends extra inference arguing with itself — plan, check, discard a bad branch — then it answers. You may only see a "thinking" pause. You are still buying those tokens.
This is a product setting, not a personality upgrade. Hosts expose "effort" or a separate reasoning model. High effort on a one-line change is a slower, pricier intern. Low effort on a multi-step outage is a confident shrug.
Use it when the obvious first answer is the dangerous one:
Skip it when you already know the file.
One honest limit: a beautiful chain can still be a hallucination. The steps look like proof. They are still generated text. Evals tell you whether the extra spend bought fewer wrong finals, or just longer wrong finals.
What this unlocks
You stop using one model setting for every ticket. Fast path for the rename. Paid extra pass for the mystery. That is how you keep a hackathon agent cheap on the easy turns and sharp on the one that would have wrecked the demo.
