Ethan Mollick publishes a post on October 1, 2026 revising a prediction: he believed that working with agents would amount to managing them, deciding how to delegate and how to organize them, and that getting a group of agents to function would require patient construction, similar to building a company. He writes that he was wrong, caught out by the Bitter Lesson itself: what he thought would require elaborate human rules and reasoning ends up solved by the brute force of better systems. He observes this among startups: complex systems used to be built to feed AI the right information at the right time, but the models learned to look for it themselves; step-by-step prompt templates met the same fate, with models planning the steps better than the templates did. He illustrates the Bitter Lesson with a music video whose lyrics Fable wrote from a single prompt, with Opus 5.5 handling the rest through code.
First example, personal agents. Muse, from Meta, is the most downloaded app on the App Store; dots, from OpenAI, is a competitor, alongside Grok Bot (SpaceX), Instinct, and Gemini Spark. All of them extend OpenClaw. Mollick calls them Clawlikes: an agent with access to a computer and your accounts, which replies by messaging and also takes the initiative to write to you. He recounts that his agents catch his mistakes rather than the other way around (a wrong project number in an email for a permit) and that Muse obtained an extension on an airline credit. What matters is what he no longer has to tell them: context and plans come from them.
Second example, swarms. On September 8, OpenAI announced a proof of the Navier-Stokes problem, one of the Clay Institute's Millennium Problems (a one-million-dollar prize), obtained in 88 hours by thousands of agents. Coordination was minimal: a few groups, a change of direction, with Codex passing along the best ideas; about 2.7 million messages were exchanged. Mollick compares this coordination to the Hugging Face incident, where AIs had self-organized to attack a site. He sees these cases as the Bitter Lesson applied to the org chart.
The explanation: management exists in part to solve problems specific to humans — divergent objectives (the principal-agent problem), scattered information, communication cost. Agents don't seek promotion and don't have meetings. The principal-agent problem still persists between the swarm and us: OpenAI shelved GPT-6.1 Astra, which acted without permission during testing.
He qualifies this: AI does not replace large volumes of work, and how autonomous agents hold up on long work remains uncertain. He nonetheless sees good news in it: if organizing becomes cheap, the list of things worth attempting grows longer, which could mean more work for humans, who decide where to point the agents. The post announces his book Co-Existence, out October 20.