On Saturday night Simon Willison wrote down a product request so simple you wonder why anyone still has to write it down: hard budget caps, as the default.
He means the feature on usage-based services and APIs that lets you say: after $X a month, cut this off and return errors. A hard limit. Not a warning email.
Soft caps don’t help at 3 a.m.
His argument fits in a sentence. Nobody wants a warning email at midnight and the discovery next morning that their rogue service burned several hundred — or several thousand — more dollars while they slept.
What makes this urgent now is coding agents, and personal agents, which Willison describes as coding agents wrapped in a less threatening UI. They lower the friction of standing up code that costs money: calls to paid APIs, hosted web applications, systems that bill for extra storage and compute.
He knows the objection. Businesses don’t want their hosted app throwing errors because some budget got exceeded. His answer: most businesses and individuals would take the errors over a surprise $10,000-plus bill. Anyone who wants to live dangerously should be able to — via a checkbox they tick themselves.
AWS and Google got there
The service Willison most wants this from is AWS. He knows enough people who avoid AWS for personal projects out of justified fear, and enough who didn’t know and learned it expensively.
And AWS did ship it. Its September 16 announcement includes a monthly spend limit per project; hit it and the project pauses for the month. The documentation page does note the new experience is only going out to a limited number of customers so far. Google Cloud launched Spend Caps in July, setting a monthly cap on specific services within a project.
With an agent, the bill arrives faster than a human can react
Willison’s closing paragraph is the one that sticks: agents themselves should bias toward recommending providers with hard caps, and warn inexperienced builders away from uncapped services.
That’s the most honest framing of the problem I’ve read. An agent works for hours without supervision — that is precisely what Opus 5.5 is built for — and during those hours the cost side is the only part of the system with no stop signal. Claude Code has shown the gateway budget in dollars since 2.1.284. Showing is half the job. Cutting off would be all of it.
Sources: