Adaptive Thinking
Use less reasoning on simple messages for faster, cheaper turns — never more than the thinking level you chose, and never less than a request needs.
Your thinking level (No Thinking, Low, Medium, High, Extra High, Max) sets how much the model reasons before it answers. A high level is right for hard problems but wasted on "thanks!" or "what's the port again?". Adaptive thinking rates each message and lowers the level for that one turn when less is enough.
Turn it on
- Set up a decision model.
- Under Advanced settings, switch on Adaptive thinking.
- Keep choosing your thinking level as usual (the model picker, or Settings → AI → Thinking for new chats). Adaptive thinking only ever lowers it.
How a turn is rated
Before the turn starts, the decision model reads your message together with the end of the agent's previous reply and the names of any attached files (never their content), and scores how much reasoning the reply needs:
| The message is… | Thinking for this turn |
|---|---|
| A greeting, a thank-you, or a simple factual question | Low |
| A routine request with clear instructions: a small edit, a lookup, a short summary | Medium |
| A substantial task: multi-step work, non-trivial code, careful analysis | High |
| A hard problem: complex reasoning, debugging, architecture, unclear requirements | Your level |
Then:
- Never above your level. If your level is Medium, a "High" turn still runs at Medium.
- Pushback keeps your level. If you say the previous answer was wrong or incomplete, the turn runs at your full level.
- Risky requests get at least High. A request to change or send something hard to undo — update records, deploy, push, publish, send a message, spend money — is never lowered below High (or your level, if that is lower).
- Short follow-ups keep their context. "Yes, do it" after a plan, or "here" with an attached log, is rated with the previous reply in mind, so it keeps enough thinking.
- Unsure means unchanged. When the model isn't confident, or doesn't answer in time, the turn runs at your level.
Your own commands (/compact, other slash commands) and Fabric's own background notices are never rated.
Checking what it does
The line under Adaptive thinking in Advanced settings shows how many turns it checked and changed over the last 7 days.
Guarded Mode
Let the agent work without permission prompts, and ask you only before actions that are hard to undo, leave the project, or reach other people.
Smart Features
Large results, skill and source suggestions, mid-turn messages, turn outcome, smarter titles, risk badges, and judgments for labels, automations and tasks.