Agents
The agent page: Overview and Build
An agent opens on Overview: its last 7 days, one row per channel and cards into Evaluate, Monitor, Build and Deploy. Build holds its instructions, toolbox, triggers and a test chat that tries unsaved changes.
Reference2 min readUpdated 31 October 2026
Open an agent from Agents. Its tabs are Overview, Build, Settings, Prompt, Knowledge, Tools, Deploy, Conversations, Versions and Cost. It opens on Overview; a new agent opens on Build. Picking the agent in the top bar's switcher shows the same Overview on Home, with Test it and Edit agent.
Overview
The last 7 days, every channel. Tiles for Requests (with conversations), Error rate, Latency p50 (and p95), Tokens, Cost (at the token prices you set) and Thumbs up. The Channels table has one row each for Widget, Chat link and API (Live) and Playground (Test): last request, requests, error rate, p50 and p95.
| Card | Shows | Button |
|---|---|---|
| Evaluate | The latest experiment and its score, datasets, evaluators scoring live replies, the top scores | Run an experiment |
| Monitor | Alerts firing on this agent, and links to Conversations, Tracing and Analytics | Create an alert |
| Build | Its tools, knowledge sources, MCP servers and guardrails | Edit and test |
| Deploy | Widget and chat link on or off, the workflow it answers with, the workflows that use it | Deploy options |
Build
| Section | What's there |
|---|---|
| Instructions | The system prompt, with a chip for its model ("No key" when it has none). Save instructions (or ⌘S / Ctrl+S) publishes them to every channel and records a version; Discard drops your edits |
| Toolbox | Built in tools and every connected app's actions, each with a switch; actions marked Changes data ask before they're turned on. Up to 30 on. Rows for Knowledge, MCP servers and Guardrails with Manage links to their tabs. See Use apps as agent tools |
| Triggers | Where it answers: the Playground and this test chat (Test), the website widget, the chat link, the API and the workflows that use it, each with a link to set it up. Answer with a workflow hands its chats to a published workflow |
| Test chat | A chat with the agent. It uses your unsaved instructions, so you can try a change before saving, and works while the agent is paused. It counts as Test (Playground) traffic, and each reply is 1 credit |
The other tabs are as before: Settings, Prompt (the same instructions on their own), Knowledge, Tools (MCP servers), Deploy, Conversations, Versions and Cost.
Still stuck?
Email us with what you tried and any request id. A person replies.