by Ilia SadykovThe usual agent demo asks you to imagine the interesting part. Watch a recording. Read a transcript. Create an account, find a workspace, configure a persona, then perhaps say hello.
What if the first step were just opening a link — and the agent answered you?
That is the change behind Forge Talk. Someone responsible for the agents signs in, chooses which ones to share, and creates an expiring invitation. A guest opens it in a browser and can type to an agent or speak to it there, without creating a Forge account. It is a more direct way to learn whether an agent is useful: have the conversation, rather than study a feature list.
The wider interest in voice is easy to see. A recent Hacker News experiment with AI calling businesses found a much more complicated reception than the phrase “voice agents are here” suggests. That experiment concerns outbound phone calls, not browser invitations; it is a reminder to make the first interaction clear and consensual, not a claim that every audience wants an unsolicited agent call.
A door, not an onboarding project
There are two different people in this experience. The admin decides which agent or agents a particular invitation can reach and when that invitation expires. The guest does not need to understand organizations, model gateways or how the agent is staffed. They open the link and choose a conversation.
That separation matters if you are showing an agent to someone who has never used Forge. You can introduce the relevant character rather than sending them into an empty dashboard and hoping they assemble the context themselves. A product walkthrough can start with a question. An interview can start with a voice. A prospect can decide whether the answer was helpful before you ask them to learn the platform.
This is not a claim that the guest has a full operator account. The invitation is deliberately narrower: it exposes selected agents and the conversation, not an admin console. It can expire, and the owner can revoke it. Revoking an invitation ends open guest calls as well as preventing new ones. A shorter path into a product should not mean a broader grant into it.
Talking is different from watching a voice demo
Inside the page, a guest can hold a voice conversation rather than join a separate meeting or install a calling app. The browser sends speech for transcription; the agent responds; its reply is spoken back. The guest can interrupt. Typed and spoken turns belong to the same guest conversation, so changing how you communicate does not require starting over.
For voice to feel like a conversation, turn-taking matters as much as the model’s answer. Krisp’s engineering discussion of voice-agent turn-taking describes how background voices and noise can confuse interruption detection. That is one reason we describe barge-in as an available interaction, not as a promise that every microphone, room and browser will behave identically. Real-device checks, especially on mobile Safari, remain to be finished.
Forge already has agents that join scheduled calls . This is a different front door. You do not need to arrange a meeting to let someone experience an agent: the invitation opens the conversation in the page. The two experiences serve different moments — one for a booked call, the other for a direct introduction.
A conversation you can follow up on
A spoken answer should not evaporate when the tab closes. Voice and chat turns are attached to the guest conversation, and the admin can watch the transcript live. That makes the invitation useful beyond the first impression: someone responsible for the agent can see what was asked and where the answer needs work, without asking the guest to reconstruct the exchange from memory.
Per-link activity is visible to the admin, too. If you share one invitation for a specific introduction, you can see whether anyone used it. This is more practical than treating a link as a mysterious demo artifact and guessing whether it landed. It is also a reason to be deliberate about what an invite covers and who receives it. Guest conversations are isolated from other guests’ conversations; the admin still needs to handle the resulting transcript responsibly.
This builds on the agent-platform infrastructure we introduced earlier : scoped guest access and agents that can work behind another product’s front door. The new part is not an API primitive. It is the moment a person gets to meet the agent without first having to become a platform user.
The honest boundary
This is a working, deployed experience, not an open invitation for arbitrary public distribution. The consent and recording language still needs its final pass, and browser/device voice QA is not complete. We are not promising a finished self-serve campaign or asking readers to publish guest links indiscriminately. Nor does this link create a PDF report or automatically email one: those are separate steps in a longer journey.
The takeaway is simpler: if the product is a conversation, let people encounter the conversation first. Remove the signup, keep the access scoped, and leave a record that can be reviewed. If you want a guided look at how an agent would speak with your audience, talk to us . Tell us which conversation you would want to begin with.