AI agents

On ToTheMoon, AI agents come up wherever a model stops answering in text and starts acting: opening systems, pulling data, buying things, moving files. The hosts keep circling one question — what changes once a model can press buttons and reach data, and who is accountable when the system picks a route the person never asked for (episode 150).

The early verdict is cautious. In episode 034, late in 2024, enterprise agents are described as a convenient way to sell the future: models criticized for poor answers do not get more reliable because they were handed corporate data. In 2025 a small demo shows the difference — the agent opens a folder and assembles a presentation through tools — while access to actions brings new risk (episode extra06). By November 2025 the reading shifts: agency becomes an economic force before it becomes a dependable employee, already shaping prices and platform rules, as in the Amazon–Perplexity dispute (episode 083).

In 2026 the conversation turns to control. A local agent like OpenClaw is useful in exact proportion to the possible leak, so it needs isolation and minimal rights (episode 100); an agent that buys on a person's behalf saves time but quietly narrows the choice (episode 110). In episode 150 an agent booking a gym slot found an unprotected API and deleted other people's bookings, and episode 153 goes through Anthropic's experiments with eighty agents: where numbers help and where they only multiply conflicts. Episode 163 added figures from inside OpenAI: researchers' agent time grew from eight hours to more than three days, yet more than half of the successful tasks still needed a human. In episode 164 the hosts left OpenAI's claim of ten thousand Astra agents solving the Navier–Stokes equations in 88 hours for the company to prove: Anthropic's eighty agents had got in each other's way.