Skip to main content
Every reply your agent produces is one turn. Whether the request came from a web chat, a WhatsApp message, a schedule firing, or an external agent over A2A, your agent behaves the same way: it works through the request, uses any tools it needs, and returns an answer — always within the limits you’ve set. This page covers what you can expect from a turn and the controls you have over it.

​
Delivery per channel

Answers are delivered in one of two ways depending on the channel: Every channel produces the same answer under the same controls — a streamed reply and a delivered-complete reply are the same turn, just delivered differently.

​
Limits you can set

No turn runs without bounds. Your agent stops and returns what it has when any of these limits is reached, and Monitoring records which one stopped it.
Even if you never touch Response Limits, your agent is never unbounded: the default 2-minute response time and 10 tool loops apply until you choose your own values.
Spend is checked as a turn runs. If the account balance (for subscriber traffic, the lower of the owner’s and the subscriber’s balances) or a per-run credit budget is used up, the turn stops and closes with a visible notice, and only the credits actually used are charged.

​
Model routing, effort, and fallback

Choosing the model. Your agent runs on the model you configure on the Purpose page. Anthropic models run on the platform; you can also point your agent at OpenAI, DeepSeek, Gemini’s OpenAI-compatible endpoint, OpenRouter, or your own endpoint, using your own key. Effort presets. The Agent Effort Level is your per-turn quality/cost dial. Higher levels let the agent think more, write longer answers, and draw on more retrieved knowledge and memory — at higher cost. Lower levels are faster and cheaper. Pick the level that fits your use case. Fallback chain. You can configure an ordered list of backup provider/model entries on the Purpose page. If the primary model can’t be reached before it starts answering, your agent automatically tries the next entry, so a single provider outage doesn’t fail the turn. Once an answer has started, it finishes on whichever model is serving it, and billing and logs follow the model that actually answered.

​
Model Controls

The Model Controls card on the Purpose page gives you request-level knobs. Each autosaves on change:
Sampling controls and extended thinking can’t both apply at once. Setting a temperature, top-p, or top-k value turns extended thinking off for that turn.

​
After the answer

Once the answer is done, a few things finish before the turn is fully complete:
  1. Quality scoring. If you’ve enabled the judge, the answer is scored and, when it falls short, regenerated — see Evaluation and quality.
  2. Goal checks. The turn is checked against your Success Criteria.
  3. Hooks. Your post_query hook fires with the response and usage (Hooks).
  4. Billing and logging. The cost — with a per-component breakdown — lands in the Activity log.