A personal agent is software that runs all the time on a machine you own, remembers you across sessions, writes its own new abilities, reaches you through the apps you already use, and carries out real tasks instead of just answering questions. That last part is the whole difference from a chatbot: a chatbot talks, an agent acts. This drop is the short history of how that arrived in 2026, a tour of what these agents can actually do, and a clear look at what they cost.
Strip away the branding and a personal agent is one small loop with five defining properties. It runs always-on, as a , on hardware its owner controls. It keeps across sessions. It writes and saves its own s. It reaches you through ordinary messaging apps rather than a dedicated interface. And it executes real tasks rather than describing how they might be done. The next tab pulls that loop apart in detail, and lets you run a scaled-down version of it yourself.
That loop is what an agent is. What makes it fraught is a single trade-off, and it runs through the entire piece. Almost every genuinely useful thing an agent does, recalling your history, acting before you ask, reaching you on whatever app you happen to be using, comes from a decision a cautious product team would have blocked. That is why the breakthrough arrived as open-source software people run themselves, where the person operating the agent and the person carrying the risk are the same person. The capability and the danger turn out to be the same property seen from two sides, and that overlap is the argument the rest of the drop builds out.
Here is the whole drop, in order. what is a personal agent lays out the five properties and the loop. lineage traces the eighty-year road from the 1945 memex through Clippy and Siri to the 2026 field, and names what each important project contributed, from OpenClaw and Hermes to the governed enterprise stacks, the minimalist forks, and Poke. philosophies maps the eight camps still arguing about what an agent should even be. what is listening and security are where the risk is: every channel an agent answers on is an inbox a stranger can write to, has no known fix, and persistence plus credentials make an agent far more dangerous than a chatbot. inbox and calendar are working toy demos you can operate yourself. And how to is a practical, MTSHow-style walkthrough for setting one up, safely. The piece does not close with a verdict, because there is not one yet; it closes where you are, at a prompt, deciding what to type.