Vision · 6 min
Call tool, not tool call: where voice agents go next
Felix Huber · 8 Oct 2026
Every new capability for agents seems to go through the same three stages. First it is a demo. Then it is a tool. Then it is infrastructure nobody thinks about. Browsing went that way, and so did code execution. We think the phone is next.
Stage one: the phone as a tool
This is where we are today. An agent hits a task that needs a call — booking, confirming, asking, rescheduling — and instead of handing it back to you, it calls. You approve the plan, the agent supervises, and you get a result. That alone closes the most common gap in what agents can finish.
The hard problems at this stage are reliability and trust. The call has to go well on a bad line with an impatient person, and you have to be comfortable letting it happen.
Stage two: the phone as a channel
Once outbound calls are reliable, the obvious next step is the other direction: an agent that answers your phone, takes the message, and only puts through the call that needs you. After that come sequences. “Call these five craftsmen, get three quotes, and book the cheapest one for next week” becomes one instruction.
Stage three: agents calling agents
As more businesses put an AI on their own phone line, more calls will be agent to agent. For a long time they will still talk, because the phone network is the one interface every business already has. Then they will negotiate a faster channel and switch. The phone becomes a handshake rather than the conversation itself.
We like to call it call tool instead of tool call: the phone becomes the universal fallback interface for agents that have no API to each other.
What has to be true
- Disclosure by default. A person should always know when they are talking to an AI.
- Consent and accountability. Every call traces back to a person who approved it, with limits enforced in code.
- No silent data. No recordings, short retention, and data handled in the EU.
In the open, and roughly in this order: more languages, inbound calls, sequences of calls with one brief, and a developer plan with usage-based pricing.