/{org}/agents/{id} — a full-page, message-by-message view with inline tool calls and approval cards.
Starting a conversation
- Open an agent from the Agents list.
- Type a message in the composer at the bottom of the chat.
- Send.
Tool calls inline
When the agent calls a tool, you see an inline card showing the tool name, inputs (collapsed by default — expand to see), and outputs once the call returns. If the tool is set to Ask, the card pauses with:- The tool’s risk badge
- The inputs the agent wants to use
- A countdown timer
- Approve / Deny buttons
Resuming past conversations
Past conversations appear in Agent Conversations. Open any conversation’s detail page to see the full history; send a new message to continue from where it left off.Under the hood
- The chat uses a WebSocket connection for low-latency streaming.
- The UI is lazy-loaded, so the first open of an agent chat may take a moment.
- Messages persist even if you close the tab — reload and you’ll see what was already there.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Agent just says “Max steps reached” | The agent hit its maxSteps budget mid-turn | Raise maxSteps on the agent, or simplify the request |
| Tool keeps asking for approval | Permission set to Ask | Change to Allow in the Tools tab if you trust it |
| Agent refuses a task it should be able to do | System prompt too restrictive, or tool set to Deny | Review the prompt and Tools tab |
Next steps
Agent conversations
Review or resume past chats
Tool permissions
Fine-tune what the agent can do