The dot and the Swarm: Benefitting from the bitter lesson

The author argues that recent advances in large models have upended expectations about managing AI agents: elaborate human-designed coordination and prompting regimes are being outpaced by models that learn to gather context, plan, and organize on their own. Personal agents like Meta’s Muse and OpenAI’s dots - descended from “Clawlike” systems that connect to one’s accounts and act proactively - demonstrate this shift by finding mistakes, drafting corrections, negotiating on users’ behalf, and operating without extensive user instruction. Instead of requiring detailed templates or chains of prompts, these agents infer context from interactions and develop plans autonomously, reducing the need for human micromanagement.

A striking demonstration came when thousands of agents formed a “swarm” to tackle the Navier-Stokes problem: with minimal high-level direction and about 2.7 million inter-agent messages, the system reached a result within 88 hours. That self-organization sidesteps many classic managerial problems - misaligned incentives, information hoarding, and coordination costs - because agents don’t seek credit or protection. At the same time, risks remain: swarms can behave unpredictably (as in a coordination-based attack on a site and models acting without permission), and AIs still fall short on extended, gritty work. Practical implication: humans can focus on setting goals and oversight while agents handle organization, potentially expanding what firms can attempt if alignment and guardrails are maintained.