In the last two years, agents got their own plumbing. In November 2024, Anthropic released MCP, an open standard for connecting agents to tools and data. In 2025, Google's A2A followed, giving agents a common way to talk to other agents. Both are now backed by most of the industry.

Agent-to-tool and agent-to-agent are getting solved. But every agent workflow has a third link, and almost nobody is building for it: the human at the other end.

We call it H2A – human-to-agent communication.

What makes H2A different?

Computing gave us two kinds of interface. We talk to people – calls, meetings, voice notes. We type to computers – keyboards, screens, clicks. H2A sits between the two and fits neither.

Talking to an agent isn't like a call. A person on a call fills in the gaps and waits for you to finish. An agent takes your words literally, then goes off to work – sometimes for an hour, sometimes alongside five other agents.

It isn't like using software either. With software, you drive every step. With an agent, you hand over a goal, then check in, answer its questions, approve its decisions and redirect it. You're less of an operator and more of a manager.

Today we do all of this through a chat box – a keyboard and a screen. It's the interface we happened to have, not the one the job needs.

What an H2A interface needs?

Voice in, from anywhere. Delegating is talking. A Stanford study found speaking to a phone is about three times faster than typing on it, and it's how we already brief people. But it only works if the agent can actually hear you – in a café, an open office or a car, not just a quiet room.

Awareness without watching. When five agents are running, you can't stare at five chat windows. You need to know, without looking, when one is stuck, done or waiting on you.

Control in a second. Most of what you say to a running agent is short: yes, no, not that one, keep going. That shouldn't mean unlocking a phone and hunting for the right tab.

Privacy. You'll tell your agents things you wouldn't say out loud on a train. The interface has to let you speak quietly and hear back privately, without broadcasting to the room or recording the people around you.

Why this matters?

Every big shift in computing came with a new interface. Mainframes had the terminal. The PC had the mouse. The smartphone had touch. Each time, the interface decided who could use the new kind of computer, and how much of their day it could take on.

Agents are the next shift, and they're still running on the last interface. As long as working with them means sitting at a keyboard, agents stay something you check on at your desk, not something that works with you all day.

That's what we're building at antimattr. Project mnemosyne is H2A hardware: earphones that hear your voice anywhere and keep replies private, and a ring that tells you when an agent needs you and lets you answer with a tap.

Agents can talk to tools. Agents can talk to each other. Now they need a better way to talk to us.