Simon Willison: Agents Need Hard Budget Caps, On by Default
The most upvoted AI post on Hacker News this weekend isn't a model launch. It's a 500-word argument for a checkbox. Simon Willison's "We're going to need default hard budget caps on pretty much everything" hit 579 points within a day.
The argument is short. Coding agents, and personal agents that are really coding agents with a friendlier face, make it nearly free to spin up code that calls paid APIs, hosts apps and bills for storage and compute. A soft cap, meaning a warning email at midnight, is useless if a runaway service burns through several thousand dollars while you sleep. What's needed is a hard limit: after $X a month, return errors. And it should be the default, with "remove the budget cap" as an opt-in checkbox that spells out who pays.
The timing is better than it looks. AWS quietly shipped project spend limits on September 16 in its new builder experience (a project gets paused for the month when it hits the limit), though only for a limited set of customers so far. Google Cloud launched Spend Caps in July. Two of the three big clouds have moved within three months, which says the agent-driven bill shock is already showing up in support queues.
The best line is the last one. Willison wants agents themselves to prefer providers with hard caps and to warn inexperienced builders away from uncapped services. That turns a billing feature into an agent-safety default. Today's agent decides which cloud to deploy to. That choice needs a cost-risk prior, and almost nobody has written one into a system prompt or a skill yet.
Link: simonwillison.net/2026/Oct/3/default-hard-budget-caps/
← Back to all articles
The argument is short. Coding agents, and personal agents that are really coding agents with a friendlier face, make it nearly free to spin up code that calls paid APIs, hosts apps and bills for storage and compute. A soft cap, meaning a warning email at midnight, is useless if a runaway service burns through several thousand dollars while you sleep. What's needed is a hard limit: after $X a month, return errors. And it should be the default, with "remove the budget cap" as an opt-in checkbox that spells out who pays.
The timing is better than it looks. AWS quietly shipped project spend limits on September 16 in its new builder experience (a project gets paused for the month when it hits the limit), though only for a limited set of customers so far. Google Cloud launched Spend Caps in July. Two of the three big clouds have moved within three months, which says the agent-driven bill shock is already showing up in support queues.
The best line is the last one. Willison wants agents themselves to prefer providers with hard caps and to warn inexperienced builders away from uncapped services. That turns a billing feature into an agent-safety default. Today's agent decides which cloud to deploy to. That choice needs a cost-risk prior, and almost nobody has written one into a system prompt or a skill yet.
Link: simonwillison.net/2026/Oct/3/default-hard-budget-caps/
Comments