Thinking

Open in product: preshos.com/agent-ops/thinking

Thinking is the visible reasoning behind agent actions: the agent’s working notes on what it is trying, what it found, and what it plans next. How hard it reasons follows the run’s reasoning effort. It runs alongside the transcript, not instead of it — actions and outcomes stay on the record either way.

Reading it well

• Read thinking when the session surprises. The reasoning usually names the wrong assumption before the action shows it.
• Thinking explains; the transcript decides. Drift gets stopped at the action, not debated in the notes.
• Long reasoning on a simple goal is itself a signal. If the agent is thinking hard about a lookup, the goal is probably underspecified.

What it is not

• Not the record. Attempted, approved or denied, changed — those live on the outcome, not in the notes.
• Not a place to paste secrets. Credentials never go in chat, reasoning, code, tests, or fixtures. Names of variables where sources name them; values never.
• Not proof of scope. Sensible-sounding reasoning about an out-of-scope system is still out of scope — question it the same way.

What next

• Sessions: the transcript-and-preview surface thinking runs inside.
• Model selector: which model the reasoning comes from.
• Approving agent actions: deciding on the action, not the rationale alone.