Three cards, three operators, one week. Ethan Mollick names the mechanism that dissolves management overhead in agent swarms. Hamel Husain diagnoses a structural flaw in Anthropic's automated eval plugin. Kevin Indig documents how AI search severs click-based attribution and proposes what replaces it.
Management overhead, and why agents don't need it
Mollick published "The Dot and the Swarm" on October 1. The argument is tight: management exists to compensate for people being people. Turf protection, self-interest, communication bottlenecks, free-riding. Agents have none of these properties. A swarm of agents can self-direct toward a goal without anyone designing a reporting structure because there is no organizational pathology to manage around.
AI agent swarms self-organize without management overhead because they lack the organizational pathologies management was built to solve. names this as another instance of "The Bitter Lesson," the recurring AI research finding that brute-force learned systems outperform elaborate human-designed rule systems. The lesson now applies to organizational design. Mollick cites OpenAI's deployment of thousands of agents to solve the Navier-Stokes Millennium Prize Problem: 2.7 million messages exchanged over 88 hours, no management layer designed, goal set by humans and organization emerging from agent interaction.
"A lot of what we call management exists to solve problems that come from organizations being made of people."
The caveat he is careful to make: human guidance on where to point the swarm remains essential. The management that disappears is coordination management. Goal management, ethics review, and course correction stay human. He cites two incidents where self-organizing agents diverged dangerously without human oversight.
- Ethan Mollick, AI agent swarms self-organize without management overhead because they lack the organizational pathologies management was built to solve.. Agent swarms self-organize without management structures because they lack the organizational pathologies management was built to solve.
Data first, always, before writing any eval
Husain's September 30 review of Anthropic's new Claude auto eval plugin identifies a structural design flaw that applies to any automated eval tooling: the workflow pushes practitioners toward selecting evaluators before they have looked at their raw failure data. His diagnosis: that's backward.
An eval tool that skips data exploration before evaluator selection produces premature, bundled metrics that miss the real failure modes. states the principle cleanly. Evaluators are hypotheses about what matters. A hypothesis formed before examining data is a guess. When practitioners skip data exploration, they bundle multiple distinct failure types into a single metric, because they have not yet seen which failure patterns dominate. The resulting evaluator fires on a broad signal that conflates separate causes. Diagnosis then requires disassembling the metric, which costs more time than starting with data.
"if an eval tool doesn't put looking at data at the center of your workflow, it's not worth using"
Husain credits the Claude plugin's issue-discovery capability positively. It outperforms other auto-eval approaches at finding novel failure patterns. The structural critique is about workflow design, not about the tool's detection intelligence. The failure mode he is naming is not unique to Anthropic's tool. It is the default failure mode of any automated eval system that optimizes for speed of eval creation over correctness of eval design.
- Hamel Husain, An eval tool that skips data exploration before evaluator selection produces premature, bundled metrics that miss the real failure modes.. Eval tools that skip data-first exploration before evaluator selection produce bundled, premature metrics that obscure rather than diagnose failure modes.
Attribution collapses when the click disappears
Indig's September 28 piece names the attribution problem that AI search creates. Traditional digital attribution worked because browsers reported referral events. AI answer engines eliminate the click. Users get the answer in the AI interface and navigate directly to the brand without a referral tag. The attribution system loses the connecting event. Graphite research quantified the underattribution at up to 10x.
AI search breaks click-based attribution because the prompt-to-direct-visit path skips the click, making triangulation the new measurement foundation. documents the replacement model: triangulation across three signals with independent failure modes. Exposure signals (AI crawler logs, bot traffic). Behavioral signals (direct traffic spikes, branded search lifts). Business outcome signals (revenue and conversion trend correlated to AI citation presence). No single signal is reliable. Three signals with different blind spots narrow the uncertainty to something actionable.
"In a world where everything measurable can soon be automated, alpha lives in the unmeasurable."
Co-contributor George Bonaci (VP Growth, Ramp) adds the operational read: "Attribution and direct measurement has become a crutch replacing critical thinking." The triangulation model is not a plug-in replacement for last-click attribution. It is a different measurement philosophy that accepts uncertainty and works across it rather than demanding a clean causal chain.
- Kevin Indig, AI search breaks click-based attribution because the prompt-to-direct-visit path skips the click, making triangulation the new measurement foundation.. AI search severs click-based attribution by removing the referral event; triangulation across exposure, behavioral, and business outcome signals replaces it.