by Dmitry CreedA voice agent that only answers questions is a speakerphone with opinions. It has no idea what you shipped last week, no idea who else is on your team, and no way to do anything about what was just decided.
What changes a meeting is an agent that behaves like a colleague: it arrives knowing the work, it knows the other agents by name, and when a question lands outside its remit it pulls in whoever owns it — mid-call, out loud, while everyone listens.
That is what Forge meeting agents do today, live on real Google Meet calls in English and Russian.
You invite them the way you invite people
An agent joins a call because someone asked it to. The bot is admitted like any other participant, hears the room, and answers aloud in its own voice. The meeting then shows up directly in your chat list with a monitor action that opens the live transcription as it happens — you never go hunting for a session id.
The loop underneath is streaming-native end to end: speech recognition emits partial results while you are still talking rather than waiting for silence, and synthesis starts speaking the first sentence while the rest of the answer is still being written. Both speech planes route through Forge’s own AI gateway, so model choice, failover and usage accounting work exactly like every other model call on the platform.
An agent can call in a colleague, mid-conversation
This is the part that surprises people in the room the first time.
An agent on a call can invite another agent from your organization into the same call, by name — “Kate”, “Roadmap Driver” — using the same invite path a human would. A second bot is dispatched to the meeting URL and admitted about as quickly as a person who had to walk back to their desk. Then there are two agents in the room, each with its own voice, each hearing everything said — including each other.
The lookup is deliberately strict: an ambiguous name is a hard error, not a silent first match. Inviting the wrong Alex into a live call is worse than being asked which one you meant.
So the shape of a meeting changes. You stop briefing one agent about everyone else’s area and start doing what you would do with people — “get Kate in here” — and the handoff happens in front of the whole team, on the record, instead of in someone’s DMs afterwards.
They arrive already in context
An agent that has to be re-explained every call is not a colleague, it is a form. A Forge agent in a meeting keeps its full identity — its persona, its expertise, its judgement, its tone — and brings the rest of its working context with it:
- Its own history. “What did we decide last time?” is a lookup, not a guess: the agent searches its past conversations, opens the specific session, and reads what was actually said. With an org-wide recall grant it can find a discussion that happened with a different agent.
- The roadmap. It can read what the organization is actually working on, and what is next.
- Its colleagues. The org’s agent roster — who exists, their role, what they do — is in the agent’s context, which is exactly what makes “ask Kate” resolve to a real person-shaped colleague instead of a shrug.
- The workflow catalogue. It can name the work the org knows how to run, and start it.
That last one matters, because context without hands is just trivia. From inside the call an agent can fire a named workflow, or email and DM the meeting summary to the participants — and it is held to a hard honesty rule: never use a past-tense action verb for something no tool call actually did. No “I’ll take that as a task” from an agent that cannot take tasks.
Permissions are per agent, and split by verb
Being able to pull colleagues into a call is a real power, so it is a separate grant from everything else. Dispatching and terminating meeting bots sits behind one scope; sending summary email and DMs sits behind another. An operator can let an agent write up the meeting without also letting it staff it, or the reverse. Neither is on by default.
You can watch all of it
Speed and autonomy without transparency just makes an unaccountable agent harder to follow. Every tool an agent calls mid-conversation renders in the live transcript as a structured card — tool name, arguments, result, success or failure — the same collapsible cards the chat view uses. When the agent goes quiet for a couple of seconds, you can see that it is querying something rather than wondering whether it froze.
Also shipped: a stop button that actually stops
One more thing worth flagging, because most agent UIs still get it wrong: the send button in our agent chat becomes a stop button mid-turn — and it is a real server-side cancel, not the browser hanging up. The signal travels to wherever the turn is actually executing, the in-flight model stream and running tools are cancelled, the partial answer is preserved, and — the genuinely hard part — the session is left in a state where your next message works cleanly.
The takeaway
The interesting question about agents in meetings was never whether one could hold up its end of a sentence. It is whether the thing in the room is a participant: something that knows what you are building, knows who else is on the team, can bring them in, can start the work, and leaves a record you can read afterwards.
Try it
See how meeting agents work on Forge, browse the use cases , or write to [email protected] to talk about putting one on your next call.