Agent Settings
Agent Settings controls the default behavior of one agent.
Agent Settings
Agent Settings controls the default behavior of one agent.
Agent scope only
Branding, team management, billing, and provider keys are organization-level settings. This page is only for one agent's AI, retrieval, escalation, and voice defaults.
Tabs In Agent Settings
| Tab | Purpose |
|---|---|
| AI Settings | Model, variant, reasoning, temperature, input tokens, conversation history, and prompt defaults |
| Voice | Voice enablement and local native voice configuration |
| Escalation | Loop handling, sentiment thresholds, and escalation keywords |
| Retrieval | Expansion, hybrid search, reranking, deduplication, and caching |
How To Use This Page
AI Settings
Use this for the agent's baseline text behavior. OpenRouter models show provider variants when available and a reasoning effort control when the model supports it. Choose Auto to use the model's default reasoning effort. Temperature is disabled for models that do not support it. Keep the prompt focused and operational.
Conversation history
Agents summarize older conversation history when a chat approaches the model's input limit, so work can continue without sending the full transcript. The AI receives the summary alongside the most recent messages, and stored history is left unchanged.
| Control | What it does |
|---|---|
| Conversation history | Turns automatic summarization on or off. |
| Compaction model | Uses the chat model, or a model you pick for summaries only. |
| Summarize at | Estimated total input that triggers summarization. |
| Summarize down to | Approximate input size to keep after summarization. |
| Keep recent | Budget for the newest complete turns kept word for word. |
How the values resolve:
- Every threshold defaults to Auto. Auto derives from the model's known input limit: summarize at 85% of it, summarize down to 70% of that value, and keep recent 30% of the target.
- "Use the same model" follows the agent's chat model and its provider credentials. A dedicated compaction model is resolved on its own and may bill differently.
- Leaving a control untouched keeps its stored value. Choosing Auto again resets only that control, and changing one control does not change the others.
- Explicit thresholds must be positive whole numbers, keep recent below summarize down to, and summarize down to below summarize at. A threshold above the model's input limit cannot run, so the turn falls back to leaving the oldest messages out of the model input instead.
What to expect:
- Summaries lose detail compared with the original messages and add model usage on top of the chat.
- Original messages stay in the conversation history under its normal retention rules. A summary replaces only the older messages it covers, and those messages are not sent to the model again.
- With conversation history off, no new summary is created and the AI may lose earlier details. An existing summary is still used, and older messages can be left out of the model input when needed to fit the context limit.
- If summarization cannot run (an unknown input limit, no eligible older messages, an unavailable model, a provider failure, or a conflicting update), the turn continues with the context it has. When the input still does not fit, the oldest messages are left out of the model input, and the chat shows a quiet status notice.
- Summarization decisions happen between model steps and never rewrite history. A published summary is reused by later steps and turns, a conflicting concurrent update falls back instead of overwriting a newer summary, and a call that has already started cannot be recalled or steered. When a run stops, the next attempt reloads the last published summary rather than replaying the messages it covers; work that was never published is lost.
- Status notices and context-management activity are content-free.
Voice
Use this when your agent supports voice experiences. Pair changes here with testing in Voice Lab.
Escalation
Use this to define when AI should hand off or ask for help sooner.
Retrieval
Use this to control how the system searches and refines source content before generation.
Retrieval Terms In Plain Language
| Setting | What it means |
|---|---|
| Query expansion | Broadens the original user question so retrieval can match close variations |
| Hybrid search | Combines semantic retrieval with keyword-style search |
| Reranking | Reorders candidate results so the strongest chunks appear first |
| Deduplication | Removes repeated or near-identical chunks before the model sees them |
| Caching | Reuses recent retrieval work to reduce repeated lookup cost and latency |
Change This Or Fix Something Else?
Change Agent Settings when the default model behavior is wrong across many conversations.
Use Sources when the knowledge itself is missing or outdated.
Use Skills when the agent needs reusable behavior or instructions across multiple situations.
Use Interceptors when a rule must always win.
Best Practices
- change one major setting group at a time
- test in Playground before relying on production traffic
- use Interceptors for hard rules instead of overloading the prompt
- review Analytics after meaningful changes
Related Docs
Playground
Test answers, retrieval, prompts, and overrides before production.
Interceptors
Rules for escalation, blocking, tagging, and guaranteed replies.
Skills
Attach reusable instructions to agents without overloading the base prompt.
Voice Lab
Test native voice sessions, local settings and backend tasks.