Thinking & Reasoning
Reasoning models work out a plan before they answer. Dashboardx Designer keeps that working-out visible, so when an agent builds the right dashboard for the wrong reason — or the wrong dashboard for an interesting reason — you can see which happened.
What you see
Above each reply that has reasoning behind it, a row of trace chips appears:
- 💭 Thinking — this reply has a reasoning trace. While the agent is still working, the chip reads Thinking… and pulses.
- One chip per sub-agent the reply delegated to — the reviewer, the analyst, and so on.
Click a chip to open the Agent Trace pane for that message. It shows:
- The Thinking section: the model's own working-out, rendered as markdown, with how long it took. It opens expanded; collapse it with one click.
- Calling: … while the agent is mid-step, so you can watch progress instead of wondering whether it hung.
- The sub-agent tree: which specialist was asked what, what it returned, and whether it is waiting on you.
The trace is per message. Open another message and the pane follows it.
When thinking appears
Thinking shows up when the model you are talking to has reasoning capability — look for the green Reasoning tag on a model in Settings → Models.
Models without it produce a single reply and there is simply nothing extra to show; the chip does not appear.
Reasoning is on by default for models that support it. There is no separate thinking switch to hunt for — pick a reasoning-capable model and you get the trace.
Why background work does not burn reasoning
Naming a conversation and compressing long history happen constantly and gain nothing from deep thinking. Those tasks go to the Global Small Model you set in Settings → Agent, so your main model's reasoning is spent on the dashboard.
Reading the trace
The trace is the agent's draft thinking, not a transcript of its actions. The authoritative record of what it actually did is the tool-step list in the same conversation — the trace explains why it chose those steps.
Useful things to look for:
- Wrong assumption about your data — visible before it reaches the canvas, while changing course is cheap.
- A plan that quietly grew — if the reasoning drifts into scope you did not ask for, say so before it lands.
- A reviewer disagreeing — expand the sub-agent chip to see what a reviewer pushed back on.
If the reasoning is not worth your attention, collapse it and read the reply. Nothing is required of you.
Long tasks and approvals
When reasoning leads into an action that changes your work, the agent stops and asks. The trace shows a pause marker while it waits for you — nothing is timing out in the background. See Settings → Permissions.