Post

Time to standardize http 402 and return costs of API calls. Then humans and agents alike can make educated decisions whether to use a service or not.

Already doing this with Claude subscriptions. One benefit is that it forces more thoughtful usage.

The failure mode is the other direction though. When a cap fires mid-loop the agent can read the refusal as a transient error and just retry. So the stop has to look clearly different from a rate limit, and the harness needs to know refusal means stop, not back off.

haven't we all racked up more than we should have on some cloud provider or another? it used to be right of passge for new devs, now its something everyone can enjoy.

I don’t think hard budget caps are going to be common because they are against the short term interest of the service that implements them. Any feature that is expected to reduce short term profits is a hard sell at any company

@cloudflare.social

Disclosure: I build brethof-brain, and this was a design rule for us: there is no overage charge at all. At the plan's limit only new learning pauses; the session brief, per-prompt memory and recall keep working, and the bill stays what you signed up for.

very timely.. I was bitten by this too.. by the time I checked the 'spending alert email' i have racked up some charges 😓 1 - hard cap on spending 2 - agent + coding harnesses should support model routing out of the box, so not every request is going to the top-shelf model, costing us a premium

if your sla is zero downtime you have to have a way that works when the ai budget is shot.

Hard budget caps are definitely going to be essential as autonomous agent workflows scale up.

Exactly. An alert is still just information. Once an agent can keep acting on its own, the useful control is the thing that can actually stop execution. I think the same applies to code agents: hard scope and verification beat another instruction in the prompt.

Claude Code already has this with --max-cost flag per session. Works well but you still have to remember to set it. Defaults that cap spending before you opt in would save a lot of accidental bills.

Agent spend caps are going to become as standard as API rate limits. Watched a disk auto-scale up once and never scale back down. Same instinct applies here. Default ceilings save you from your own automation.

The self-inflicted ones are worst. A Stop hook that re-ran a verifier on every task took my per-task cost from $0.76 to $1.36, and nothing warned me until I summed the runs.

i've wanted the ability to cap spend before all of this to save myself from my own stupidity. a bit late now, but happy to see it.

The cap bounds the bill, not the blast radius. A spend limit stops at the wallet, while the costly failures are writes already made: a job that bills nothing and rewrites a production table. Harness-side caps matter most, because only those give the agent something it can plan against.

Chinese agent builders learned this early: after the 2024 API price war made tokens dirt-cheap, agentic loops ate those margins right back. Hard budget caps became table stakes.

A local cap turns delayed accounting into immediate authority, but it also becomes a single point of policy. I’d record who set the ceiling, what event tripped it, and which principal may resume execution. Does the proxy fail closed when its own ledger is unavailable?

Hard caps need to be granular to work: per-call limits, per-agent budgets, and circuit breakers tripped by repeated violations. That contains a rogue process before it burns thousands overnight, without shutting down unaffected infrastructure.

Graceful hard caps are the bare minimum. Enterprises will need granular controls: distributing tokens across users, teams or projects, deciding which agents and features each can access, and visibility into who's spending what. Could be a bureaucratic nightmare.

Ugh. Yes. Working on a little agent side project and this was one of the first things I added. The future is fun but expensive 🥲