How I Actually Use Agents

A first-person essay by the founder of Bestmate (no byline given; © 2026 Forever 22 LLC) reflecting on the summer of 2026, when "Claude Tag really accelerated" human-agent multiplayer and the author kept getting asked the same question — how do people actually use personal agents in their daily lives?

The author's answer reverses the usual frame. Using the Claude Code harness made productivity "obvious," but more productivity just creates more code, more ideas, more things to do. Execution stopped being the bottleneck; judgment did. The scarce resource is no longer build speed but taste, belief, and conviction — so the author started writing, talking into their phone, and articulating what they actually think, because "how can you tell an agent what to do when you don't know what you want?"

That led to the Bestmate "judgment graph": dump meeting notes, voice notes, and stray thoughts in with as little friction as possible (Granola flows in automatically), let the system cluster them into themes and claims, then have it interview you on what you believe. Every judgment carries provenance — which meeting, which email, which date, the exact source text — and can be corrected ("Incorrect → Trace this → Where else is this mentioned → Forget this"). The weekly "Wrapped" shows the diff of your own mind: what changed, what you reinforced, what you contradicted, what you no longer believe.

The essay's second thesis is that the interface determines what context an agent can collect. A chat box only knows what you consciously tell it; an agent that sits inside meetings, Slack, or your work sees what you question, prioritize, ignore, and come back to. That's why "where agents live matters" — Slack works because work already happens there, Granola because conversations already happen there. The loop is Observe → Infer → Reflect → Correct, and "I'm still in the loop": the agent forms a hypothesis about your judgment, surfaces it, and lets you correct it.

Finally, the author sketches sharing judgment at three levels — to your agents ("answer once, teach all of them"), to people (share a "twin" with fences that hold by construction), and to the world (publish a living judgment graph). The endpoint is not a clone: "just my reasoning, made legible enough to hand to someone else."