Smart Features
Large results, skill and source suggestions, mid-turn messages, turn outcome, smarter titles, risk badges, and judgments for labels, automations and tasks.
Besides Guarded mode and adaptive thinking, the decision model can make these small judgments for Fabric. Each has its own switch under Settings → AI → Decision model (Jev) → Advanced settings, with a tooltip saying exactly what the model is asked and a line showing its activity over the last 7 days.
Agent
Large tool results
When a tool result is too big for the conversation (a long log, a big search, a long document), Fabric saves it to a file. With this switch on, the decision model picks the parts the agent's request needs — the matching messages of a Slack search, the failing lines of a log, one section of a document — and the agent gets those parts plus the saved file in about half a second. Without it (or when nothing matches, or nearly everything does) a small model writes a summary, which takes several seconds.
Every source call states a one-line purpose, which is what the parts are matched against.
Skill and source suggestions
Before a message is handled, it is compared with your skills and with sources that are set up but not active in the session. When one clearly fits — your calendar source for "find a free slot on Saturday" — the agent is pointed to it. Nothing is enabled on its behalf.
Conversations
Mid-turn messages
When you send a message while the agent is working, the decision model judges whether it corrects the running task (delivered right away, so the agent changes course) or is a separate follow-up (queued for after the turn). A message you split over two sends ("when you create the event later," … "call it Standup") is merged into one request.
Turn outcome
After each turn, the agent's last message is read to judge whether it finished, needs your input, or got stuck. A session that is waiting for you — a question, a decision, a blocker — moves to Needs Review, so it stands out. In tasks, a subtask that asks for input or gives up counts as failed.
Smarter titles
A session that starts with small talk ("hey 👋") isn't titled until you ask for something, and every few messages the title is refreshed if the conversation has moved on. A title you set yourself is never replaced.
Permissions
Risk badges
Each permission prompt is tagged with what the action does: deletes, sends, publishes, credentials, system, spends. Works in every permission mode.
Guarded mode
Offers the Guarded permission mode.
Automations and tasks
| Switch | What it does |
|---|---|
| Semantic label rules | Label auto-rules can be yes/no questions ("Is this about billing?") instead of patterns. See semantic rules. |
| Automation conditions | An automation can carry a condition written as a question ("Is the user reporting a bug?"), checked before it runs. See semantic conditions. |
| Task verdicts | When a task's verifier forgets its VERDICT: line, its reply is read as pass, fail or unclear instead of asking again. Inferred verdicts show their confidence in the task editor. |
| Targeted task repairs | After a failed check that doesn't name the broken steps, only the steps the failure points at are re-run. |