AI Agents Won't Save You

Walled gardens will kill the AI agent delusion

The world has gone mad for agents. Not James Bond or The Man From U.N.C.L.E. AI agents, as in autonomous bits of software that are meant to traverse the web and do various things on your, or other agents’, behalf. It’s a seductive idea: the web is a substrate and large language models are the enabling technology that allow autonomous entities to traverse it while you sit on your couch watching the Lakers.

And it’s doomed to fail.

There’s a structural problem lurking beneath the agentic enthusiasm: agents only work in open ecosystems, and the modern internet is anything but open.

This agent can’t help you. Image sourced from here.

The Fallacy of Frictionless Orchestration

The bull case for AI agents assumes a world of seamless interoperability: your agent can book a trip through Expedia, adjust your calendar on Google, generate slides in Notion, and message your team in Slack, all without you lifting a finger. That Lakers game isn’t going to watch itself.

But that vision only works if those platforms cooperate. In practice, they don’t.

Every major tech incumbent has spent the past decade closing its ecosystem, not opening it. Google throttled free API access to YouTube and Maps. X locks down its developer endpoints. Meta’s data firehose dried up after the Cambridge Analytica scandal. Apple treats privacy as both moral posture and competitive moat. The age of move fast and build on top of us is over.

That means today’s agents are like diplomats without passports. They’re credentialed and powerful on paper, but they are unable to cross borders.