TLDRocket
Sign in

Ben's session #2

Ben's Bites

Personal agents aren’t magic — mostly folders, files, and instructions. That’s why they can feel different without being a new product at all.

Based on reporting by Ben's Bites — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Ben’s latest post makes a simple point: the hype around personal agents is mostly packaging. OpenClaw, Hermes, Grok Bot and the rest may look like new creatures, but under the hood they’re built from the same ingredients as other agents — instructions, tools and context. The difference is how those pieces are organised, and how much control the user gets over them.

That organisation starts with files. An agent points at a folder, reads its instruction file, and works with whatever context lives there. If the session gets too long and compaction wipes it, the important stuff can be saved in a text file instead of being lost. Ben frames that as the real foundation of a “personal agent”: not a special model, just a setup that keeps state outside the chat.

He splits the idea into two paths. One is a single main agent, Jarvis-style, with task-specific files it can read when needed. The other is several task-specific agents, each with its own instruction file, personality and job description. Grok Bot fits that second pattern, with separate threads for separate agents. But Ben says the same thing can be done with pinned threads in Codex or Claude, aimed at different folders.

Memory, in this setup, is just another file. It is a log of what happened, not some magical internal recollection. Some agents can have their own memory files, and they can also share memory by reading one another’s files. The same goes for tools. Codex and Claude Code can use a computer-use tool to navigate, click, install, fill forms and make files. Grok Bot uses a shared computer where each chat gets its own screen.

The last piece is automation. Agents can be scheduled, triggered, or taught a workflow, and Ben points to Grok Bot’s “teach” feature as an easier way to record a task and turn it into a routine. His bottom line is blunt: if the instructions and tools are there, agents can do a lot — book flights, order food, negotiate quotes, organise files, collect invoices, create decks, build websites and apps. The trick is the setup, not the label.

My take — AI-written commentary, not fact-checked reporting

This is the part the AI crowd keeps trying to inflate into sorcery, and it really isn’t. Files, folders, permissions, schedules: that’s the boring machinery, and boring machinery is usually where the product actually lives. The winners here will be the systems that make control obvious instead of hiding it behind shiny agent branding and a very confident thread.

Read more about this at: Ben's Bites

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.