Live speaking agent

Send a Live agent into Google Meet. It listens to the room and talks only after a human invites it. When participation tools are available, it may raise its real Meet hand before asking to speak. You control it by speaking in the meeting, not from dashboard buttons.

Before the call

  1. Open Agents and create an agent with its one-time credential. Follow the Agents guide to connect your AI over MCP.
  2. Join a Google Meet from Meetings with Send to meeting and choose your agent, or run /harvest https://meet.google.com/xxx-xxxx-xxx from the hosted client.
  3. Admit the agent if Google Meet asks. The host may need to let it in from the lobby, same as any guest.
Live participation is different from the silent recorder. Choose your connected agent, not Harvest recorder, when you want a speaking participant.

How participation works

Once the agent is in the call, you control it by speaking. There are no dashboard buttons to raise its hand, grant the floor, or make it talk.

Listen-only (default)

The agent hears the room but stays silent. It does not interrupt human conversation or speak on its own initiative without permission.

Human invitation grants speech

A human must explicitly invite the agent to speak, for example by name. That invitation is the permission to give one bounded answer. After the answer the agent returns to listening. Another invitation is required for the next turn.

On current Live deployments, a human invitation alone is enough to grant speech. The agent does not need to raise its hand first when participation tools are unavailable.

Raised hand (when tools are available)

When participation tools are registered, the agent may raise its real Google Meet hand before it wants to contribute. Other participants see the hand icon on its video tile. The agent stays muted and waits. A raised hand is a request for attention, not permission to speak. A human invitation is still required.

Quiet stop

Ask the agent to be quiet during a call, for example “Harvest, be quiet.” Audible output stops immediately, any queued speech is cleared, and any raised hand is lowered. The agent keeps hearing the room but will not speak again until a fresh human invitation. It does not speak an acknowledgment when going quiet.

Listen-only vs the floor

  • Listen-only: the agent hears but cannot speak. This is the default, and also applies after a quiet stop until someone invites it again.
  • Requesting the floor (when tools available): the agent raises its Meet hand and waits. Output stays closed until a human invitation.
  • Has the floor: a fresh human invitation grants one answer. When playback finishes, the agent yields and returns to listen-only.

What you see on the Meet tile

While the agent is in the call, its participant tile may show these in-call states:

  • Listening when it hears the room and is not speaking or preparing a response.
  • Thinking when it is preparing an invited response.
  • Hand icon when it has raised its Meet hand and is waiting for an invitation. This is Google Meet's hand overlay, not a separate Harvest label.
  • Speaking when it has permission and audio is playing. Mouth movement and captions follow the actual speech.
These states appear only while the agent is in the call. Joining, processing, and finished labels on recordings in Library are post-meeting playback states. They are not shown on the live Meet participant tile.

After the call

Open Library for the recording, transcript, and summary once processing finishes. See the Library guide.

Tips

  • Invite the agent by name so it knows the floor was granted to it.
  • Human-to-human discussion stays uninterrupted unless the agent raises its hand.
  • If several agents are in one call, Harvest keeps one speaker at a time. Tell each agent to answer only when addressed by its own name.
  • For setup and MCP permissions, see the Agents guide and FAQ.

Something looks off? Email support@tryharvest.ai.