Most "AI agents" are still just chatbots
Open one of today's popular AI tools and you get a window you paste into. You copy an email in, ask for a reply, copy the reply back out, then send it yourself. The model never touched your Mail app. You are still the one moving data between the model and your software, which means you are still doing the job. A chat box that gives advice is useful, but it is not an agent. An agent acts.
A real agent works inside your apps, not in a tab
The difference shows up the moment you ask for something that spans more than one app. "Reply to Sarah's last email and tell her the deck is ready, then drop the link in the design channel." A real agent opens Mail, writes the reply, sends it, switches to Slack, finds the channel, and posts. You watch it happen. It uses the same apps you use, with the permissions you grant, instead of asking you to be the courier.
It remembers, so you stop re-explaining yourself
Today's chatbots forget everything the moment you close the tab. You re-explain your job, your projects, and the people you work with every single session. A real agent keeps a memory. Tell it once that your sister is named Anna and her birthday is March 3, and in late February it can remind you and draft the card. Tell it your standup is at 9:30 and it stops asking. Memory is what turns a clever tool into something that actually knows you, and it should live on your machine, not in a company database.
It does work while you are away
The most useful thing an assistant can do is hand you finished work before you ask. A real agent runs on its own schedule. While you sleep it reads the overnight email, drafts the three replies that obviously need sending, and leaves them waiting for one look and a nod in the morning. At 8am it gives you a short brief: what changed, what needs you, what can wait. You did not prompt any of it. That is the line between a tool you operate and an agent that works for you.
It runs on your Mac, not on someone's server
An agent with access to your email, messages, and files is the last thing you want streaming to a server you cannot see. The honest version runs on-device. Your audio, your memory, and your documents stay on the Mac. That also makes it fast, because there is no round-trip to a data center for every action, and it keeps working on a plane with no wifi. On-device is not a privacy nicety. For an agent this deep in your life, it is the only responsible default.
Voice is the interface that finally fits
Typing instructions to an agent is often slower than doing the task yourself, which defeats the point. Speaking is faster than typing for most people, and it works hands-free while you are reading, cooking, or walking. Hold a key, say what you want, release. Good voice control answers in under half a second and works in every app, including the ones with no API and no plugin. Voice is not a gimmick bolted on top. It is the input that makes an agent worth reaching for instead of just doing the task by hand.
You should be able to read its source code
An agent that can read your inbox and act in your name is asking for an enormous amount of trust. The only way to earn that is to let anyone read the code. Open source means you, or someone you trust, can check exactly what it does, what it stores, and what it sends. Closed agents ask you to take their word for it. For something with this much access, "trust us" is not good enough.
This is what we are building
Jarvis started as a dictation app that over 100,000 Mac users rely on. Jarvis 2.0 is the agent version: it works across your apps, remembers what you tell it, runs routines while you are away, takes voice as the primary input, and runs on-device. It is open source, so you can read every line. It launches in summer 2026, and early access goes to the waitlist first.