The decide Tool
Let the agent classify, route, score and yes/no-check up to 200 items at once, with probabilities instead of prose.
With Agent tool (decide) switched on, new sessions give the agent a decide tool. Instead of asking a full language model to write an answer, the agent asks the decision model typed questions about text or JSON and gets probabilities back in a fraction of a second.
When the agent uses it
| Task | Tool |
|---|---|
| Sort 150 emails into 6 folders | decide: one choice question over a batch of items |
| Is each ticket about billing? | decide: a yes/no question |
| Rate support replies on a 4-level rubric | decide: a score question |
| Summarize, extract fields, draft text, explain | call_llm, or the agent itself: a decision model can't write text |
Rule of thumb: if the answer is one of a fixed set of options, a level, or yes/no, decide fits. If the answer is words, it doesn't.
You can simply ask: "Use decide to sort these 120 issues into bug, feature request and question."
Question types
- choice: pick one of 2–255 options. Returns the most likely option, a confidence and the full probability distribution.
- score: an ordered rubric of 2–10 levels, lowest first. Returns the expected level and per-level probabilities.
- noul: a yes/no probability.
Up to 20 questions per call, over a batch of up to 200 items. Results keep the input order; a failed item carries an error while the rest stay usable.
Reading the results
- A confidence below 0.5 means the model is unsure; the agent reports that, asks you, or falls back to a full model rather than acting on a coin flip.
- An answer is never permission: the agent re-checks before acting on a classification (for example, that a chosen folder exists).
Limits
- Text or JSON only, up to about 96 KB per item (about 48 KB with Laya). Images are not supported yet.
- Roughly 0.05–0.5 s per call; batches run 8 items at a time.
Smart Features
Large results, skill and source suggestions, mid-turn messages, turn outcome, smarter titles, risk badges, and judgments for labels, automations and tasks.
Sources
Connect Fabric Agents to external data — MCP servers, REST APIs, and local folders. How sources work, how to add one, and how the agent uses them in conversations.