Agents
Creating & configuring an agent
Agent settings, budgets, and the danger zone.
Creating an agent
From the dashboard, click + New agent and give it a name — that's the only required field. You're dropped straight into the new agent, and every other setting can be configured afterward from its tabs.
Each agent card on your dashboard also has a live/paused toggle above it, letting you take an agent offline across every channel without deleting it.
Settings tab
- Name — shown internally and, depending on the channel, to customers.
- Model — picks both the model and the underlying provider in one choice. See Choosing a model.
- System prompt — tone and behavior instructions. Hidden when the agent is connected to an external Foundry agent (see below). See Writing a system prompt.
- Temperature — a 0–1.5 slider controlling how deterministic vs. varied responses are. Lower is more consistent; higher is more exploratory.
- Daily token budget — an optional cap (prompt + completion tokens, reset daily at UTC midnight) across every channel the agent is published to. Leave blank or 0 for unlimited. Once exceeded, the agent stops answering until the next day — a guardrail against runaway usage or abuse.
- Retrieval settings — grounded-answers toggle, chunk count, candidate count, and rerank threshold. See Retrieval & grounded answers.
- Escalation & ticketing — lets the model hand off to a human or mark a conversation resolved. See Escalation to a human.
Connecting to an external Foundry agent
If your organization has an Azure AI Foundry project connected (under Organization settings → AI providers), you can point an agent's Model picker at "External agent — Microsoft Azure AI Foundry" instead of a hosted model. This hides the system prompt and retrieval settings here — those live in Foundry — and asks for the project endpoint, agent name, agent version, and optionally an API version. The agent authenticates using your organization's Azure service principal.
Deleting an agent
The Danger zone at the bottom of Settings deletes the agent along with its knowledge, tools, and channels. This is irreversible and requires confirming in a dialog — there's no recovery after deleting.