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 manage the conversation's context. These are the runtime levers; changing one here changes the agent's character across every task it handles.
Reach this page from Agent Config → Agent Behaviour in the left navigation. The page reads "Control how your agent behaves, communicates, and handles tasks."
The page is organized into two sections: Core Controls (the day-to-day levers) and Advanced Controls (model selection, context, and safety toggles).
Core Controls
Autonomy
A slider with markers Low, Med, High: "How independently the agent works." The default is Med.
Higher autonomy means the agent commits to multi-step plans on its own; lower autonomy means it confirms more steps before proceeding.
Agent Capability
A dropdown: "Controls how deeply the agent thinks." The default is Auto.
Auto
The platform decides based on the input. Marked Not Supported on an executor version that doesn't support it.
Intelligent Planner
—
Next Best Step
The Planning Model field is hidden with this option.
Fast
"Fast strategy strips persona, FAQ, CoT validation, and learning cases for lower latency." Available only on executor versions that support it.
While Response Mode is Flash, Agent Capability can't be changed.
Response Mode
A dropdown: "Speed vs quality of responses." The options are Advanced, Balanced, and Flash. The default is Advanced.
Advanced produces higher-quality responses at a slower rate; faster modes return more quickly with shallower processing. The Planning Model and Execution Model fields appear only in Advanced. A mode that isn't available for the account is flagged, and the page can't be saved until you choose another.
Memory settings have moved
User persona memory and FAQ memory are no longer on this page. Both are under Agent Config → Reinforcement Learning (Preview), are off on a new agent, and aren't used while Agent Capability is Fast. See Reinforcement Learnings.
User Persona tab — Enable User Persona. When on, the agent remembers what an individual user has asked about, preferred, or corrected, across conversations. The tab lists stored persona memories, and you can delete one.
FAQ memory
FAQ tab, under Agent Config → Reinforcement Learning — Enable FAQ. When on, the agent can use stored FAQs and their canonical answers when a question matches a known FAQ. The tab lists the stored FAQs.
Advanced Controls
Model Family
A dropdown that filters the Planning Model and Execution Model lists to one provider family: "Filter available models by provider family." It lists the families AI Gateway serves for your organization. Agents can use Google and OpenAI. A family marked Dev only isn't available in production. Clear the selection with the × beside it.
If the saved family is no longer available in your AI Gateway, the page says so, and you choose another family before saving.
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. It appears when Response Mode is Advanced and Agent Capability is Auto, Intelligent Planner, or Fast.
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).
Execution Model
A dropdown selecting the model that the agent uses to execute tasks. It appears when Response Mode is Advanced. The options are the models AI Gateway can serve for your organization, so the list differs between organizations and changes as models are added — open the dropdown to see yours. A model the gateway can't serve is greyed out. A saved model that's no longer available stays in the field, marked Unavailable. A model the agent's executor version doesn't support is marked Not Supported, and the page can't be saved until you choose another.
Below the model fields, a banner reads: "Model output may differ across cloud providers (e.g. Azure, AWS, GCP) due to hosting and infrastructure differences. Verify outputs for critical use cases."
CoT Reasoning (beta)
A toggle marked Deprecated. It can't be turned on.
On an agent that never had it on, the toggle is disabled: "Chain of Thought is deprecated and can no longer be enabled."
On an agent that already has it on, you can turn it off: "Chain of Thought is deprecated. You can turn it off, but once disabled it cannot be re-enabled." While it is on, Select Model for CoT sets the model used for CoT reasoning.
With Agent Capability set to Fast, the toggle is marked Not Supported.
Auto-compaction
A toggle: "Summarize the conversation when it nears the context window." When on, Compaction threshold sets when compaction happens: "Compact when context usage reaches this share of the window," from 50% to 95% in steps of 5. Auto-compaction appears only when the agent's executor version supports it.
Prompt Sanitization
A toggle: "Enable detection of malicious user prompts." On an agent whose executor version doesn't support it, the toggle is off, unavailable, and marked Not Supported, and it is saved as off.
Fields and obligation
Autonomy
Mandatory
Default Med.
Agent Capability
Mandatory
Default Auto. Locked while Response Mode is Flash.
Response Mode
Mandatory
Advanced, Balanced, or Flash. Default Advanced.
Model Family
Optional
Filters the model lists. Google and OpenAI in production.
Planning Model
Conditional
Advanced mode with Auto, Intelligent Planner, or Fast. Same pool as Execution Model.
Execution Model
Conditional
Advanced mode. Options and default vary by organization.
CoT Reasoning (beta)
Deprecated
Can only be turned off, on agents that already have it on.
Auto-compaction
Optional toggle
Shown when the executor version supports it. Threshold 50–95%.
Prompt Sanitization
Conditional
Shown as Not Supported when the executor version doesn't support it.
Saving
The Save button activates with unsaved changes. After saving, test the new behavior with Try Agent.
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. The Draft Version popover is where you snapshot a behavior configuration that's working, so you can recover it later.
Maintained by Abhishek Paul
AI-assisted, human-approved
Last updated
Was this helpful?