Put a ceiling on agent spending
Coding and personal agents make it easier to launch software that calls paid services, but developer Simon Willison argues that those services need default hard budget caps. A hard cap cuts off usage after a set monthly spend; an alert can let charges continue while nobody is watching.
For an agent project, check every paid API, hosted app, storage service and compute provider before launch. If a service offers only alerts, decide whether you can enforce a ceiling elsewhere. A warning email is not a spending control.
Give the coding agent a definition of done
Design engineer Lauri Lännenmäki reports building 15 website templates with Claude Code in about a week. He chose each concept and supplied reference images, typography, spacing and colors; the coding agent did most of the implementation.
His reusable move was defining done before shipping. The checklist covered six screen widths, keyboard navigation, visible focus styles, text contrast, reduced-motion behavior and serious accessibility issues. Lännenmäki still reviewed every template visually and made the final call.
Each template also includes an AGENTS.md file that tells future AI assistants where things live and which rules must not break. It is project documentation designed to guide the next edit instead of merely recording the last one.
Also
- AI assistants move into text messages TechCrunch surveys assistants for calendars, family coordination, travel and work, with availability ranging from public services to private betas.
- Plan the interface before generating it Dishant Sharma demonstrates researching an app, drafting text wireframes and feeding one phase at a time into Stitch, a tool that generates interface designs; he says the output still needs refinement.
- Keep a person behind the document Splice CEO Kakul Srivastava argues that AI-written documents can weaken workplace conversations when readers cannot tell whether a person stands behind the words.
Pass it on