AI Pulse · Under the Radar

A founder with 44,000 GitHub stars on his project doesn’t have a code editor on his machine. Matt Van Horn sends work to an AI agent via Telegram. The agent has its own email address. It runs tasks while he’s out. He feeds it full meeting transcripts, not summaries he wrote himself.

That detail is what makes the story stick. It’s not a product demo or a thought piece about the future. It’s a concrete description of how one person actually delegates to software that acts more like a junior team member than a search box.

The workflow is simple. Van Horn describes a task in plain language. The agent picks it up, works through it, and reports back. If it needs clarification, it asks. If it hits a wall, it says so. He doesn’t sit and prompt it in real time. He checks in when he’s ready.

This is the kind of thing we build into an AI command centre, where agents handle repeating work and surface what needs a human decision. The difference between a chatbot and an agent is whether you have to babysit it. Van Horn’s setup suggests he doesn’t.

What matters here isn’t the specific tools he uses. It’s the mental model. He treats the agent like a person he hired remotely. He gives it context, not instructions. He lets it figure out the steps. That only works if the agent can actually do multi-step work without constant supervision.

This is a culture artifact, not a feature announcement. It’s spreading because it’s legible. People can picture it. They can imagine their own version. The question isn’t whether this is possible. It clearly is. The question is whether your work has enough repeating structure that an agent could take a chunk of it off your plate. If you’re still summarising meetings yourself before you act on them, that’s a place to start.