Simple agents

Ben’s Bites has a simple take on personal agents: they’re mostly folders, files, and tools wired into whatever agentic system you prefer. The newsletter says a poll on the idea pulled in 996 votes, then leans hard into the same point — the magic is less magic than structure. A memory file can make an agent act remembered, but under the hood it’s still just reading text at the start of a session.

That framing matters because it cuts through the current fog around agent products. If the agent can read a folder of context, it can work with different tools, whether that’s Claude, ChatGPT Work, or something else entirely. The point is the setup, not the logo on the front end. And if the latest writing style from one model annoys you, the folder can simply be pointed somewhere else.

The post also teases a walkthrough of a clean personal agent setup in the next issue, after the author admitted their own folder got messy enough to need a desk-cleaning moment first. There’s a neat line tucked into their agent instructions too: questions should be treated as requests for an answer, not an excuse to change a bunch of things before replying. That’s the sort of small rule that makes these systems feel less like novelty demos and more like usable tools.

Around that core idea, the newsletter is packed with AI product churn. Meta AI shipped a ChatGPT-like Mac app, Berd by Block mixed folders, characters, boards, checklists, and sticky notes, and Ramp launched router.com to pick models automatically while claiming lower costs for the same outputs. Slack is pushing code and agent work directly inside Slack, which may be the most normal-sounding weird product idea in the bunch. The through line is obvious: everyone is trying to turn messy work into something the model can actually hold onto.