Advanced Agent Settings
Most teams never need this tab - the defaults are tuned to work well, and a strong personality description does most of the shaping. Reach for the Advanced tab once you have watched a few real conversations and want to nudge a specific behavior. Everything here is optional, and anything you leave blank falls back to the platform default.
Open it from Agents > your agent > the Advanced tab.
The agent Advanced tab showing stopping conditions, response tuning, and custom instruction fields
Stopping Conditions
Your agent can take several steps in a single turn - search your knowledge base, call a tool, then write its answer. Stopping conditions cap how far it goes, to protect speed and cost:
- Max steps (per turn) - the hard cap on tool-call rounds (platform default: 50). Raise it if your agent gives up before finishing multi-step tasks; lower it for faster, simpler replies.
- Timeout (seconds) - end a turn if it runs longer than this.
- Max loops - stop the agent if it repeats the same action too many times.
If you are not sure, leave these blank - the defaults suit normal support.
LLM Tuning
These apply to every turn:
- Temperature - how creative the wording is, from 0 (deterministic) to 2 (creative); the default 0.7 is balanced.
- Max output tokens - the cap on how long a reply can be (default 4096). Raise it if answers get cut off; lower it to keep replies brief.
- Extended thinking - lets the model show its reasoning on complex decisions (Anthropic models only).
For support, lean toward a lower temperature. Customers want a correct, consistent answer far more than a creative one.
Two-Layer Prompt
The most powerful setting here splits your agent's prompt into two layers:
- Identity prompt - who the agent is: personality, tone, voice. If you set this, it replaces the main system prompt, so use it only when you want full control of the framing.
- Domain instructions - what the agent does: workflow, rules, and escalation triggers (for example: "Confirm the order number before discussing a refund; if a claim is over $150, escalate to a human"). This is added as its own section on top of everything else.
Most teams only need Domain instructions for one or two firm rules - leave Identity prompt blank unless you want to rewrite the agent's core framing. A See example patterns link offers ready-made starting points.
These settings change how your agent behaves, so test after editing. Use the preview panel to run a few real questions - including ones your agent should refuse or escalate - and confirm it still responds the way you expect before you publish.
What's Next
With your agent configured the way you want, it is time to put it on your site. Head to Embed Widget to publish and install the chat widget.
Was this page helpful?