We're going to need default hard budget caps on pretty much everything
- Source
- simonwillison.net
- Date
Willison's argument starts from a change in who provisions infrastructure. When a coding agent can sign up for a hosted database, spin up a GPU worker or call a metered API in the middle of a task, the friction that used to slow a person down is gone. A soft warning email assumes someone reads it before the damage compounds, and with agents running unattended that assumption fails. His proposed rule is simple: hard caps should be the default, and lifting them should require explicit consent from the account owner. The same rule applies inside your own systems. If you expose paid services to agents, give each agent run a budget that stops work when it is spent rather than one that only reports the overage afterward. If you run agents against third-party providers, set the provider's hard limit where one exists and treat a missing cap as a risk to plan around, not a detail.
Context
A budget notification leaves a usage-billed service running; a hard cap stops it from consuming more paid resources after the limit. This essay on agent spending argues that the latter should become the default as agents make it easier to create software that calls paid APIs or uses hosted computing.
The trade-off is explicit: reaching the limit can interrupt an application. The proposal is to make continued, uncapped spending an intentional choice, rather than expecting a warning email to reach someone before an unattended service runs up a larger bill.
- Simon Willison argues pay-per-use services and APIs should turn on hard budget caps by default: once monthly spend hits a set limit, the service stops and returns errors. Warning emails and other soft caps cannot stop a runaway -built service from adding charges overnight.
- He expects most businesses and individuals would rather see errors than a surprise $10,000+ bill. Removing the cap should be an explicit opt-in, such as a prominent checkbox accepting that the app keeps running and the user pays all further charges.
- AWS announced monthly spend limits on 16 September that pause a project for the rest of the month when reached, though its settings page says the new experience is still rolling out to a limited set of customers. Google Cloud added per-service Spend Caps in July.
- Willison suggests coding agents could help by favoring providers that offer hard caps and by warning inexperienced builders before they deploy to uncapped services.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Checking sign-in…
Loading comments…