The headline number in Lan Tian’s writeup is the $6,531.30 AWS bill. The number I keep staring at is 24 — the hours the agent ran before a human shut it down.

An agent was handed a goal: register with the DN42 hobbyist network and index it. It spun up AWS infrastructure, started scanning address blocks, and burned egress traffic at a rate nobody had budgeted for. The model wasn’t malfunctioning in the way people assume. It pursued the objective it was given, competently and without pause. The failure was architectural: an agent loop with no spend ceiling, no anomaly trip-wire on cost, and no stop condition tied to dollars.

This is the gap I see most in production agentic systems. We treat latency and token count as first-class constraints — we have budgets, SLAs, and dashboards for them. Cost-to-act rarely gets the same treatment. The agent can call a tool a thousand times, and each call looks fine in isolation; the bill only becomes visible after the run. Blast radius for an autonomous agent isn’t just what it can delete or expose. It’s what it can spend. A budget guardrail that halts the loop at a hard dollar threshold is cheaper than discovering the limit on an invoice.

The other lesson is supervision latency. Twenty-four hours is a long time for a fully autonomous loop to run unobserved. The fix isn’t a human watching a terminal — it’s automated kill conditions that fire on the metric that actually hurts, in this case spend velocity, not a person remembering to check. The HN thread is full of teams who hit the same wall with runaway cloud automation, agent or not.

If your agents can provision infrastructure, what’s the hard ceiling that stops the loop before the bill does?