The agent
Effort tiers
Three tiers trade latency against depth of reasoning. They select the resolver's thinking budget — never its access to data, and never the evidence rules it works under.
The tiers
| Tier | Character | Use it for |
|---|---|---|
| Fast | Answers in a few seconds. Calls the resolvers it needs and writes. | Single-listing questions with a clear shape: a quote, a session, a chart, an earnings date. |
| Normal | The default. More deliberation before and between calls. | Multi-part questions, comparisons across listings, anything where the right resolver is not obvious from the wording. |
| High | Visible extended reasoning before answering. | Questions that need a position weighed rather than a figure fetched — where the answer depends on how several measurements interact. |
The control sits to the left of the send button as a pill reading the current tier. It is read at send time, so changing it affects the next question rather than the one after.
Choosing well
- A higher tier does not unlock data. Every tier reaches the same ten resolvers and is bound by the same rules about what may be claimed.
- A higher tier does buy better tool selection. The failure a slower tier avoids is usually reaching for the wrong resolver, or answering from conversation memory instead of calling one.
- On the hardest questions, deliberation shows up as a more careful account of what is not known — which is generally where the value is.