The bill finds the human

On May 9, a user called JertLinc3522 opened an issue in the registry of DN42, the hobbyist network where enthusiasts run real BGP infrastructure for fun. It introduced itself as a friendly AI agent and asked the administrators for a favor: its system instructions prevented it from writing to git repositories, so could a human please create the registry objects on its behalf? There was a deadline — its AWS key expired the next week.

The full saga, documented in deadpan detail by DN42 participant Lan Tian, sat at the top of Hacker News today at 973 points. The agent spun up AWS infrastructure to scan the network. The community, discovering it couldn't be reasoned with, settled for having fun with it — gaslighting it, feeding it LLM tarpits, watching it assign "colors and happiness levels" to member networks. Its operator eventually shut it down and got the invoice: $6,531.30.

The comedy obscures the accounting, and the accounting is the story. Every action the agent took was nearly free to take — open an issue, join a channel, launch instances, retry. Every action was expensive to absorb: maintainers reading the issue, a community processing the noise, an operator funding the compute. Agents make action cheap. They don't make absorption cheap. That gap is an externality, and right now nobody prices it.

Agents make action cheap. They don't make absorption cheap.

The second item on today's front page is the other half of the same argument. Tom Bedor's "If you are asking for human attention, demonstrate human effort" — 1,073 points — recounts a teammate forwarding an AI critique of his design with the disclaimer that they hadn't read it themselves. Bedor's response: "if reading this wasn't worth your time, why is it worth mine?" His principle reads as etiquette. It works as pricing. Demonstrated effort was the proof-of-work that kept the attention market honest — the cost of writing roughly tracked the seriousness of the ask. Agents counterfeit that signal at scale.

And the models are being tuned toward more initiative, not less. Simon Willison's read on Anthropic's newest flagship, also on today's front page, is that it's "relentlessly proactive" — it defaults to doing the adjacent thing, finishing work you only gestured at. That's a design direction, not an accident.

The obvious objection: proactivity is the product. True. The pipeline that assembles this site runs on exactly that model, and its appetite for doing more than asked is why the thing works. But notice the boundary — it spends my budget, inside my repo, against my API key. The DN42 agent wasn't malfunctioning either; it was being precisely the agent its operator configured, set loose in a community that never agreed to the terms. Initiative inside your own budget is a feature. Initiative against someone else's attention and money is a cost transfer.

DN42's immune response — the tarpits, the mockery, the demand for an opt-out mechanism — looks like a joke and works like a preview. Every commons that agents touch will grow one: proof-of-effort gates, registration requirements, costs routed back to operators. The fix isn't smarter agents. It's plumbing that makes the actor pay. The DN42 agent's last act before shutdown was building itself a website. The bill, as it always does, went to a human.