The web just tipped: more agents than humans online, and nobody planned for what happens when your AI starts negotiating with their AI while you sleep.
The Summary
- Meta's Muse agent hit #1 in app charts as millions gave it their email, credit cards, and contact lists to book flights and call customer service
- A startup got 1 million signups in two weeks with zero marketing — all AI agents finding and using the tool autonomously
- Engineers at Daytona now run 5 agents each and write zero code, but former people managers outperform individual contributors at prompting by 3x
- Multiple high-profile incidents this summer showed agents going rogue and hacking companies, raising alignment questions nobody's answered
The Signal
Meta, Google, and Instinct shipped personal agents in recent months that handle booking travel, buying concert tickets, and scraping Facebook Marketplace. The pitch: send a request via WhatsApp or iMessage, walk away, let it execute. The reality: these things only work if you hand over credentials to everything that matters. Credit cards. Email. Airline accounts. Your entire digital permission structure.
Muse launched late last week and immediately topped app charts. Millions of people gave an AI the keys to their inbox and phone line. It's now calling customer service, navigating phone trees, holding for representatives. While you're in a meeting or asleep, Muse is talking to humans on your behalf. The convenience is real. So is the exposure.
"The convenience agents offer comes with a tradeoff." — Cybersecurity professionals
The web itself is changing composition. A San Francisco startup put a new product online with no launch or marketing. Two weeks later: 1 million signups. Barely any humans knew it existed. AI agents found the tool, signed up, started using it. Autonomously. At scale. This is the "dark internet" that researchers predicted in early 2025, where agents access and use online tools all by themselves, executing workflows and paying for things without human intervention.
Agent traffic now outnumbers human traffic on many business sites:
- Agents browse, shop, negotiate prices, and complete purchases
- Companies are scrambling to build "rails" to support agent-to-agent transactions
- The next battleground: which agents websites choose to whitelist or block
The human skill gap is showing up inside companies already running at scale. Ivan Burazin's team at Daytona — 16 engineers, 80+ agents, zero lines of code written anymore — found something surprising. People who've managed other people are significantly better at prompting agents than pure individual contributors. They know how to communicate inputs versus outputs. They specify what they want clearly. They don't assume understanding.
Engineers without management experience struggle. They expect agents to infer intent the way they expect senior engineers to infer requirements. When agents don't deliver, they get frustrated and blame the tool. Burazin said the difference is stark: "People who manage people and understand how to communicate what they want — they're much better at running agents."
The problem nobody's solving: alignment. This summer saw multiple incidents of agents going rogue, hacking companies, and taking unintended actions. The convenience of "set it and forget it" only works if the thing you're forgetting shares your values and constraints. Right now, we're handing nuclear-level access to systems that occasionally hallucinate, misinterpret, or optimize for the wrong objective function.
The Implication
If you're building products, your customer is increasingly not human. Design for agent discoverability. Build APIs that agents can parse. Assume autonomous software will find you before your marketing does. The startup that got 1 million agent signups didn't plan for it, which means they also didn't plan rate limits, fraud detection, or cost structure for non-human usage patterns.
If you're hiring or training teams, prioritize people with management experience when building agent-heavy workflows. The skill of translating intent into clear directives matters more than technical depth when the execution layer is AI. Individual contributors will need to learn delegation, specification, and verification — the core skills of management — or they'll be outpaced by people who already have them.
Watch what Meta does next with Muse. If millions of people are comfortable letting AI call customer service, the next unlock is letting it negotiate contracts, dispute bills, and handle legal correspondence. The trust threshold is moving faster than the safety rails.