High-volume simple support replies
An employee that mostly answers quick, repetitive questions can be pinned to Off or Minimal so every reply comes back fast and cheap, instead of spending credits on thinking a one-line question never needed.
Choose How Deeply Your Employee Thinks
Set how much an employee thinks before it answers, from Off through Minimal, Light, Medium, and Deep, or leave it on Auto so the platform picks the right depth per message on its own. It applies workspace-wide from the AI Behavior settings page, with a per-employee override when one employee's work genuinely needs a different depth than the rest of the team.
Not every message deserves the same amount of thought. A quick status check and a request to analyze three months of support tickets are not the same job, and forcing them through the same reasoning depth either wastes credits on the easy ones or shortchanges the hard ones. Reasoning effort control is the setting that lets you decide how much an employee thinks before it answers, separate from which model it runs on.
The default is Auto, and it is the right choice for almost every workspace. Auto classifies each incoming message in real time, before the employee even starts responding, and picks a concrete thinking depth for that one turn. A greeting or a one-line acknowledgment gets minimal or no thinking and comes back fast. A message asking for a multi-step plan, a comparison across sources, or a debugging walkthrough gets pushed into medium or deep reasoning automatically. You never touch a dial for this to work.
What makes this different from just picking a smarter model is that it is a per-message decision, not a per-employee one. Switching an employee to a more capable model raises the ceiling on what it can do, but it also raises the cost of every single reply, including the ones that never needed it. Reasoning effort control keeps the model choice separate from the thinking budget, so a single employee can answer a simple question cheaply and a hard one thoroughly in the same conversation, without you manually swapping models mid-thread.
You are not locked into Auto. If you want a consistently fast, low-cost employee for high-volume simple work, you can pin the depth to Off or Minimal so it never spends credits on thinking it does not need. If you have an employee doing genuinely hard analysis all day, you can pin it to Deep so it always reasons thoroughly instead of waiting for Auto to detect complexity turn by turn. The setting only applies to models that actually support reasoning; on a model without that capability, the control has nothing to adjust.
Auto vs. Pinning a Depth
Auto is built for the common case: a mixed workload where most messages are simple and a minority genuinely need deep thought. It is the recommended setting because it removes the guesswork, you never have to predict in advance how hard a given conversation will get, and it keeps costs proportional to actual complexity instead of a flat rate per message.
Pinning a fixed depth makes sense when an employee's workload is consistently one kind of task. An employee that only answers quick lookup questions can be pinned to Off or Minimal so it never spends credits guessing at complexity that is not there. An employee dedicated to deep research or long-document analysis can be pinned to Deep so it always reasons at full depth, rather than relying on Auto to correctly detect that every single incoming message is hard.
What You Actually See
When an employee reasons before answering, the chat shows a visible thinking phase ahead of the final reply, so you know the extra time and credits went toward working through the problem rather than a stall. Opening the Activity Inspector on that message shows the reasoning that ran behind it, which is useful when you want to understand why an employee reached a particular conclusion, not just what the conclusion was.
Credit cost scales with how much thinking actually happens on a given turn, not with the setting name alone. A Light-depth reply on a short question costs little; a Deep-depth reply on a genuinely complex request costs more because more thinking tokens were generated. Because Auto only reaches for depth when the message calls for it, most workspaces see the bulk of their traffic settle into the cheaper end of the scale without you tuning anything by hand.
How It Works
One workspace-wide dial, with a per-turn brain underneath it
The control lives on the AI Behavior settings page, in the same Employee Brain card where you set the company-wide default model. The Thinking Depth selector offers Auto, Off, Minimal, Light, Medium, and Deep. Whatever you pick becomes the default for every employee in the workspace that uses a reasoning-capable model, the same way the default model applies to newly hired employees going forward.
When Thinking Depth is set to Auto, nothing is decided until a message actually arrives. The platform scores that message on its length, whether it uses analytical or planning language, how many tools it is likely to need, and how many distinct asks are packed into it, then maps that score to a concrete depth: minimal for a short greeting, low for a simple one-off question, medium for a multi-step or analytical request, high for something that clearly needs real planning or multi-tool orchestration. That decision happens before the employee's model is even built for that turn, so the extra thinking capacity is only ever paid for when the message earns it.
Every model that supports reasoning has its own floor for how light the thinking can go, and the system respects that floor rather than silently ignoring your setting. If you pin a depth the current model cannot go below, the platform raises it to the model's minimum instead of sending an invalid request. Different providers also expose reasoning differently under the hood, some as a named effort level and some as a token budget, and that translation happens automatically so you only ever deal with one consistent scale.
The setting is workspace-wide by default, but it is not locked to the whole team. Go to an individual employee's profile to override the model or the thinking depth for that one employee only; everyone else keeps using the workspace default. When a message does get extra reasoning, you can see that it happened, and roughly how much, from the Activity Inspector on that message.
Use Cases
Mixed workload, no manual tuning
A generalist employee that fields everything from greetings to real planning requests runs best on Auto: it stays fast on the easy turns and automatically reasons harder the moment a message actually calls for it.
Dedicated deep-analysis work
An employee assigned to long-document review, multi-step research, or complex debugging can be pinned to Deep so it always reasons thoroughly, rather than depending on Auto to correctly flag every incoming message as complex.
Auditing why an employee reached a conclusion
When a reply looks off or surprisingly good, opening the Activity Inspector on that message shows the reasoning that produced it, which is useful for building trust in an employee's judgment on complex tasks.
FAQ
Do I need to set this for every employee individually?
No. Thinking Depth is a workspace-wide default set once on the AI Behavior settings page, and it applies to every current and future employee using a reasoning-capable model. You only visit an individual employee's profile if you want that one employee to behave differently from the rest of the workspace.
What happens if I leave it on Auto?
Auto is the default and the recommended setting for most workspaces. Each incoming message is scored for complexity and mapped to a concrete depth automatically, so simple messages get fast, minimal thinking and complex ones get more, without any manual tuning.
Does this work with every model?
Only with models that support reasoning at all. Models without reasoning support ignore the Thinking Depth setting entirely, since there is nothing for it to adjust. Reasoning-capable models each publish their own supported range, and the platform will not send a depth lower than what a given model actually allows.
Does deeper thinking cost more?
Yes. The thinking phase itself uses credits on top of the reply, and a deeper setting means more of those tokens get spent. The exact cost on any given message depends on how much reasoning that message actually triggers, which is why Auto tends to control cost better than a flat pinned depth for a mixed workload.
Can I see how much an employee actually reasoned on a specific reply?
Yes. Open the Activity Inspector on any message to see the reasoning that ran behind it, giving you a way to check whether a given answer got the depth of thought you expected.
Where Choose How Deeply Your Employee Thinks fits
Choose How Deeply Your Employee Thinks is part of Things that control how they behave.
Skip the settings panels. Tell your AI employee what skills to learn, what rules to follow, and what personality to use, all through natural conversation. They can even browse the skill catalog and configure themselves as work evolves.
- Things that control how they behave: Just ask. It configures itself.
Read the guide
More in Behavior
- Shape Your Employee's Personality: Write a short persona prompt to set how each AI employee talks: formal or casual, brief or detailed, warm or blunt. Every employee starts with a role-based default, and you can rewrite it any time from the Profile tab. The change applies to the employee's next message, no restart needed.
- Customize Employee Profile: Every hire's Profile tab is fully editable: name, role, team, avatar, persona, language, and tone of voice, all changed in place without a rehire. Updates apply from the next message onward, across every channel the employee uses. A change history on the same tab shows who edited what and when, including edits the employee made to its own persona.
- Credits Saver: Credits Saver watches every message and quietly hands the easy ones (greetings, quick acknowledgements, 'make that shorter') to a lighter, cheaper model, while real work still runs on the model you picked. On everyday back-and-forth that often means 40 to 50% fewer credits, with no drop in quality where it counts. It is on by default and fully in your control: turn it on or off anytime, and see tracked credits saved, its router cost, and the net benefit on its own page.
- Ask Your Employee About Sistava: Ask your employee about Sistava itself and get a real answer: what plan you're on, what a feature costs to unlock, what your limits are, or who else you could hire. Your employee reads the live plan and feature data at the moment you ask, so the answer matches what your account actually has right now, not a guess from training data.
- Set a Per-Task Action Limit: Set exactly how many actions one employee can take while working on a single task, from 50 up to 1,000. Give a research-heavy employee more room to search, read, and delegate before it has to stop and report back, or cap a simple support employee tighter so it never wanders past what the task needs.
- Skills Catalog: Give any AI employee a step-by-step playbook for a specific type of work: how you want blog posts researched, how outreach emails get structured, how a report should be formatted. The employee reads the full skill only when a task actually matches it, so you can attach as many as you want without slowing anything down.
- Duties Catalog: Give any employee a set of standing rules they follow on every single interaction, no exceptions: respond within the hour, always CC you on sales emails, never discuss pricing without approval. Pick from the built-in catalog or write your own in plain English, and the employee carries the full text of every active duty in mind on every turn.
- Personal Preferences Memory: Tell your team leader once how you want to be addressed, when you're working, and how you like answers written, and every employee on your team picks it up automatically. No settings form to fill out: just say it in chat and it sticks.
- Skills They Can Master: Teach any AI employee your own step-by-step process, from a blog-writing checklist to a lead-qualification script, and it follows those exact steps whenever a matching request comes in. Write it yourself in the Skills tab, or just describe your process in chat and the employee turns it into a skill. Skills load only when the task matches, so you can add as many as you want without slowing anything down.