Skip to main content

Chat and tasks

There are two ways to run an agent from the console: Chat for a conversation, and Start a task for one job. Both are on every agent card and on the agent's Explore agent page. Your conversations and runs are private to you.

Chat​

  1. Open the Chat tab and pick an Agent, or select Chat on an agent card.
  2. Start from Explore an idea, Make a plan or Create something, or type your own message. If the agent has required inputs, fill them in under Set up this conversation.
  3. Choose how the agent uses tools and how tool actions are approved, then Send message.
SettingOptions
Tool useLet agent decide, Require tools, No tools
ApprovalAsk each time, Auto-approve this turn

Conversations lists your recent chats with the selected agent (from your latest 50 runs), with search. They are only visible to you. If the agent publishes a new version, start a new chat to continue.

The header shows the agent's engine and model, its Tools, and View agent.

Start a task​

Start a task opens a panel for one job:

  1. Fill in Task for your agent and any inputs the agent asks for.
  2. Select Start run.
  3. Review proposed tool actions before they run, unless the agent is set to run tools without waiting.

Earlier runs lists your previous runs of the agent with their outcome. Open one to see the conversation and results, or Start another task.

Explore agent​

Explore agent shows how an agent is put together before you run it:

  • Who can run it: the owner and how the agent is shared with you.
  • Agent: status (for example Live), whether it runs in Chat and Tasks, and how many tools are allowlisted.
  • What it is made of: Model (and whether it is your provider), Tools (with Exact allowlist), Inputs, Output, Engine (with its turn limit) and Human approval (for example Every tool call waits for you).

Below the map, Tools lists every allowlisted tool and its connection. Explore usage opens the agent on the Usage tab.

Tool approvals​

By default, every tool action waits for the person running the agent. An owner can turn on Run tool actions without waiting for approval when sharing the agent or connecting it to Slack. Either way, tools your MCP Gateway marks as needing confirmation still pause the run, and time spent waiting counts toward the agent's run timeout.

On this page