Agent Behavior
Agent Behaviour controls how the agent thinks and answers — not what it knows or what it can do, but how it goes about its work. The settings on this page choose the model that runs the agent, decide how independently the agent operates, balance response speed against response quality, and toggle reasoning features that produce visible chains of thought. These are the runtime levers; changing one here changes the agent's character across every task it handles.
Manage Agent → Agent Behaviour
Reach this page from Manage Agent → Agent Behaviour in the left navigation.
The page is organized into two sections: Core Controls (the day-to-day levers) and Advanced Controls (model selection and reasoning toggles).
Core Controls
Autonomy
A slider with markers Low, Med, High. The default is Med.
Autonomy controls how independently the agent acts without checking back with the user. Higher autonomy means the agent will commit to multi-step plans on its own; lower autonomy means it confirms more steps before proceeding. Use High for well-scoped automations the agent has been thoroughly tested on; use Low for new agents whose behavior you want to keep visible.
Agent Capability
A dropdown. Default is Auto.
Controls how deeply the agent thinks about each request. Auto lets the platform decide based on the input. Other settings range from quick lookups to deeper reasoning. Auto is the right default for general-purpose agents; pick a specific setting if you've measured that Auto under-thinks or over-thinks for your workload.
Response Mode
A dropdown. Default is Advanced.
Controls the speed-versus-quality trade-off in responses. Advanced produces higher-quality responses at a slower rate; faster modes return more quickly with shallower processing. If your users wait on streaming output, choose a faster mode; if your agent produces a final document users review, Advanced is the right default.
Memory
A toggle. Default: off.
When on, the agent remembers user preferences across conversations — what an individual user has previously asked about, expressed preferences for, or corrected. When off, every conversation starts fresh. Turn this on for assistant-style agents that benefit from personalization; leave it off for agents handling discrete, stateless tasks.
FAQ memory
A toggle. Default: on.
When on, the agent can use stored FAQs and their canonical answers when a question matches a known FAQ. When off, the agent answers each question from scratch even when it has been asked before. Leaving FAQ memory on is the right default for support-style agents.
Advanced Controls
Execution Model
A dropdown selecting the model that the agent uses to execute tasks. Default in this engagement: Gemini 3 Flash. Observed options include GPT 5.1, GPT 5.2, GPT 5 Nano, Gemini 3.1 Pro, Gemini 3 Flash, GPT 5.4, GPT 4.1. The exact list may vary by tenant and over time.
Heavier models produce more thorough reasoning at higher cost per call; faster models cost less and respond more quickly. Pick the lightest model that meets your quality bar; switching to a heavier model is a one-click upgrade if you need it later.
Planning Model
A dropdown selecting the model that the agent uses for planning tasks — deciding what to do, in what order, given an incoming request. The option list is the same as Execution Model. Default in this engagement: Gemini 3 Flash.
The planning model can differ from the execution model. A common pattern is to use a heavier model for planning (because planning happens once per request) and a lighter model for execution (because the executor may run many tool calls per request).
CoT Reasoning (beta)
A toggle. Default: off.
When on, the agent generates a visible chain of thought as part of its response — the reasoning steps it took to reach the answer, alongside the answer itself. Useful when you want users (or auditors) to see how the agent arrived at a conclusion, particularly for high-stakes or compliance-sensitive tasks. The feature is marked beta.
Enable Prompt Sanitization
A toggle, currently shown as off with a Not Supported pill.
When supported, this enables detection of malicious user prompts before they reach the model — prompt injection attempts, jailbreak prompts, content designed to subvert the agent's instructions. The pill indicates the feature is unavailable in the current configuration. The exact condition that controls availability is not surfaced in the UI.
Note — Depth gap The condition that gates Enable Prompt Sanitization (a model that doesn't support it, a tenant setting, a region restriction) was not determined during capture. If you need this feature, contact your platform administrator to confirm what's required to enable it.
Fields and obligation
Autonomy
Mandatory
Default Med.
Agent Capability
Mandatory
Default Auto.
Response Mode
Mandatory
Default Advanced.
Memory
Optional toggle
Default off.
FAQ memory
Optional toggle
Default on.
Execution Model
Mandatory
Default varies by tenant; Gemini 3 Flash in this engagement.
Planning Model
Mandatory
Same option pool as Execution Model.
CoT Reasoning (beta)
Optional toggle
Default off.
Enable Prompt Sanitization
Conditional
Saving
The Save button activates with unsaved changes. After saving, test the new behavior in the Agent Preview pane — model changes in particular can shift response style noticeably, and behaviors that worked under one model may need re-tuning under another.
Why this matters
Behavior settings affect the agent's character — how it reasons, how it phrases, how confidently it acts. The same instructions executed by Gemini 3 Flash and by GPT 5.4 produce different agents, even when neither configuration has changed otherwise. Treat behavior settings as production parameters you tune carefully and version, not as defaults you set once and forget. The Draft Version popover is where you snapshot a behavior configuration that's working, so you can recover it if a later experiment goes sideways.
Last updated