LLM Settings
By the end of this page you will know how to set your agent’s reasoning strategy, its execution limits, how its reasoning appears to users, and which model and guardrails it uses.
Prerequisites
- An AI Agent open in the editor (Create an AI Agent).
- Admin or Developer role.
The LLM Settings tab has three sections: Executor Configuration (with UX Settings), LLM Configuration, and LLM Policy.
Executor Configuration
This controls how the agent thinks and how hard it works on a request.
Choose a reasoning mode
Pick how the agent reasons:
| Mode | How it works | Good for |
|---|---|---|
| ReAct Pattern | The agent thinks and acts in interleaved steps. | Dynamic tasks. |
| Planner → Executor | The agent plans all steps first, then executes them. | Complex workflows. |
Set Max Steps
Max Steps caps how many reasoning steps the agent takes per request (1–50). Higher values allow more complex, multi-step work at the cost of latency and tokens.
Set Max Tool Calls
Max Tool Calls caps how many tools the agent may call per request (1–100).
Start conservative — a handful of steps and tool calls is enough for most assistants. Raise the limits only if the agent runs out of budget before finishing genuinely complex tasks.
UX Settings
UX Settings control what end users see while the agent reasons and uses tools. Each has a Default (auto) option that gives builders full detail in the Studio preview and a cleaner experience in production.
Extended Thinking
Set Extended Thinking to Default (off), On, or Off. When on, the model can “think through” complex problems before responding — better reasoning, more tokens. Turning it on reveals a Thinking Budget (tokens) slider (1,024–32,768; default 8,192).
Reasoning Display
Reasoning Display controls what appears in the thinking block:
- Full — reasoning text plus tool chips.
- Summary — tool chips only.
- Hidden — no thinking block.
Default is auto: Full in the Studio preview, Summary in production.
Tool Display
Tool Display controls how tool calls appear:
- Technical — raw tool names.
- Friendly — user-facing labels (and de-duplicated).
- Hidden — no tool chips.
Activity Status
Activity Status shows a brief summary such as “Checking details…” when the agent starts using tools. Default is on.
LLM Configuration
The model the agent runs on.
Choose a model (or use the project default)
LLM Configuration is optional. Leave it empty to use the project’s default model. Override it only when this agent needs a different model than others in the project — use the model selector to pick a specific system or project LLM configuration.
An LLM configuration defines the provider, model, and settings such as temperature. Configurations are managed at the project level. Selecting one here shows “LLM Config selected - save to apply” — it takes effect when you Save.
LLM Policy
An LLM Policy applies guardrails — PII handling, topic limits and content filtering. See LLM Policies for what a policy can control and how to build one.
Add a policy
In the LLM Policy section, click Add Policy. Choose an existing active policy for the project, or click Create New Policy to open the policy wizard.
Review the applied policy
Once applied, the policy card shows its status and a summary of what it enforces — PII action, allowed and blocked topics, a count of blocked-content rules, and active-content removal. Use Edit to change it or Remove to detach it. See Apply a policy to an agent.
Removing a policy detaches it from the agent — the policy itself is not deleted and can be re-applied later.
Save your changes
Click Save in the header. Model and policy selections show a “save to apply” toast and are only persisted on Save.
Common issues
- The agent stops before finishing — It hit Max Steps or Max Tool Calls. Raise the limits, or simplify the task.
- End users see raw tool names — Set Tool Display to Friendly and give tools user labels (see Attach tools).
- No policies to choose from — The project has no active policies. Create one from the Add Policy dialog.