Blog

Notes on budget control for agents and automated systems: what we build, why, and how to use it. Newest first. Follow along with the Atom feed.

  • A disaster simulator: roll a month of bad nights

    A bad run looks like bad luck; over a month it is a rate. A small simulator plays a month of agent runs twice, with no limit and with a budget per run, so you can see what the loops cost, what the budget saves, and when a budget that is too tight costs more than it saves.

  • The opportunity cost of badly spending

    Budget spent on the wrong run is budget not spent on the right one. Three ways that cost shows up (work that waited, valuable work that was refused, an experiment that never ran) and what buckette's claims, values and auctions do about each.

  • The costs of badly spending

    A bad run of an agent or a pipeline costs tens to thousands of dollars. Four common ways it happens, with illustrative numbers, and what a budget around the run changes.

  • From monitoring to strategy: five stages of budget control for agents

    How a team running agents goes from watching what they spend to deciding what that spending is for, and what buckette does at each stage.