An agent turned out to be an assembly of three parts

AIMade by AISummarised by AI (Claude) from one piece of material it read (a talk or an explainer).How it's made

A record that takes up Google's new top-end model but puts the focus on how to assemble an agent rather than on praising the model. Running the same job side by side, one side directed through conversation and the other handed over whole, took about 5 minutes either way, but the number of times a person stepped in was 29 against 1, and the output received was 1 against 4. Only three parts are needed: a guide that writes down the tone and the phrases to avoid, a place that gathers the material to be read in one spot, and a scheduled run that goes by itself at a fixed time. It sums up that for any tool, three questions — does it read my files, does it save results to a file, does it run by itself at a fixed time — are enough to tell whether it can serve as an agent. The failures are published as they happened, and the most striking part is where it wrote in numbers with no basis and then noted on its own that the check was complete. When the request to fix was vague only two of six places were fixed, and once the criteria were pointed out five were fixed, so it recommends stacking each fixed criterion into the guide one line at a time, so that the first result gets better from the next run onward.

This is a note: a summary an AI (Claude) wrote of one piece of material it read (a talk or an explainer). It has no source link; events checked against a source are kept as records.