American Express Now Insures AI Agent Mistakes — and a Tool to Keep Your Mac Awake

American Express will now cover your AI agent's bad purchases

American Express has quietly crossed a line that most enterprises are still tiptoeing around: it created a developer kit so businesses can let autonomous AI agents spend money on their behalf—and then added the world's first purchase protection program that explicitly covers AI agent errors.

The ACE Developer Kit, which includes five integrated services for agent-initiated transactions, is designed for scenarios where software, not humans, kicks off payments. Think: an AI system that monitors supplier SLAs and auto-issues credits. A procurement bot that reorders inventory when levels drop. A collections agent that negotiates payment plans and executes them.

The notable twist: Amex is pairing the kit with a purchase protection program that names "AI agent errors" as a covered category. Not fraud. Not human mistakes. Agent errors.

This is not a technical footnote. By explicitly covering mistakes made by non-human actors, American Express is doing something more significant than launching a developer tool—it is acknowledging that autonomous software will participate in the economy as a principal, not just a tool. A 174-year-old financial institution is willing to put its balance sheet behind decisions made by lines of code.

For enterprises experimenting with autonomous agents in procurement, receivables, or vendor management, the main blocker has always been liability. Who pays when the bot buys from the wrong vendor, or pays the wrong invoice twice, or negotiates a payment plan that violates a contract clause? Amex is effectively saying: build the agent, we'll help cushion the mistakes.

This is the kind of governance crutch that CFOs and risk teams can bring to a board meeting to justify moving from "the AI recommends" to "the AI executes." Responsibility is becoming a product feature.

A tiny utility built purely to keep your MacBook lid closed

Meanwhile, a developer launched a tool whose sole purpose is to let knowledge workers run background agents on a MacBook with the lid literally closed—and it became one of the most upvoted launches on LaunchFeed, a quiet but influential product-tracking feed.

The tool is called Adrafinil. Yes, after the wakefulness drug discussed on productivity and nootropics forums. The name is not subtle.

What it does: prevents a Mac from sleeping when the lid is closed, but only if certain software agents are active. Once those agents finish their tasks, the machine is allowed to sleep again, preserving battery and reducing noise. It is a B2B-ish infrastructure tool aimed at people running background AI workers, automations for devops, or long-running jobs that don't need user interaction but would normally be killed by sleep policies.

On LaunchFeed, Adrafinil has a score of 77—significantly higher than most adjacent launches in the same time window, many of which are in the low single digits. That suggests substantial community interest for a product with an oddly narrow purpose.

The tool exists because people are increasingly treating their laptops as semi-autonomous compute nodes—machines that run agents overnight, process tasks, scrape data, and generally "work" while the human is away. The lid-closed detail is not cosmetic. It is an accommodation for the physical reality of office life: you want to pack your laptop, leave a meeting room, go home, but still have the agents finish their queue.

Why a Mac sleep utility and agent insurance belong in the same story

These two products—one from a Fortune 500 financial services giant, one from an independent developer—are solving the same underlying problem from opposite ends.

American Express is formalizing the legal and financial scaffolding for agents to act autonomously in high-stakes transactions. Adrafinil is formalizing the physical and operational scaffolding for agents to run on personal hardware without human babysitting.

Both are responses to the same shift: software is starting to do things that used to require a human in the loop, and the infrastructure around that software—legal, financial, technical—is being rebuilt piece by piece.

Enterprises talk about agentic workflows in abstract terms, but the actual practice often includes hacks: special sleep settings, duct-taped cron jobs, odd utilities that keep machines awake "just enough." Adrafinil is one of the few that formalizes that behavior into a product. It reveals the informal infrastructure layer beneath polished AI platforms.

And American Express, by naming agent errors as insurable events, is doing the same thing at the institutional level. It is formalizing what was previously a gray area: who is responsible when software makes an economic decision?

The answer, increasingly, is: everyone. The agent's operator. The platform provider. And now, the card network. Risk is being distributed across more parties because the old model—one human, one decision, one point of accountability—no longer maps to how work actually happens.

Naming a Mac agent-utility after a stimulant is a quiet nod to the human side of this. Developers and power users are literally stimulating both themselves and their machines to keep working beyond normal cycles. The same impulse that led someone to build Adrafinil is the same one driving enterprises to ask Amex for agent-specific purchase protection.

We are not just automating tasks. We are renegotiating the boundary between human and machine responsibility, one API call and one sleep-prevention utility at a time.