Meta’s chief AI officer says AI agent swarm outperformed 100 engineers

A swarm of AI agents built by Meta outperformed a team of 100 human engineers on specific tasks, according to the company’s Chief AI Officer Alexandr Wang. The revelation came during a conversation with Y Combinator president Garry Tan at YC’s Startup School 2026, where Wang laid out Meta’s philosophy on autonomous AI systems and the unglamorous plumbing that makes them work.

Markdown files and cron jobs: the anti-hype stack

Wang’s description of Meta’s agentic infrastructure reads less like a science fiction screenplay and more like a competent DevOps setup from 2018. The agents use persistent memory stored in markdown files. They’re scheduled through cron jobs, the same basic task-scheduling utility that’s been running on Unix systems since the 1970s.

The philosophy is deliberately simple and modular. Rather than building some monolithic AI brain that tries to do everything, Meta’s approach breaks tasks into discrete loops where agents can evaluate their own output, correct course, and iterate. What made the system outperform 100 engineers wasn’t raw intelligence. It was the combination of robust evaluation methods, the ability to run continuously, and a feedback architecture that let agents improve without human intervention. Wang emphasized that the right evaluation system is the critical variable, not the sophistication of the underlying model.

From Scale AI to Meta’s AI chief

Wang founded Scale AI, the data labeling and AI infrastructure company. Meta acquired Scale AI for $14.3 billion, and Wang joined as Chief AI Officer in June 2025. Wang brought the evaluation-first mindset from Scale AI, where the entire business model revolved around making AI systems measurable and accountable.

The YC event’s programming leaned heavily into themes of agentic loops and self-improving systems, with Y Combinator developing internal solutions like the QM multi-agent harness and promoting programming geared towards closed-loop systems.

Muse: Meta’s consumer-facing agent play

Meta has been developing Muse, internally known as Hatch, a personal AI agent designed to handle tasks across multiple applications, including email management, scheduling, and payments, all coordinated by a single agent that maintains context across different tools. Muse is scheduled for public release in September 2026.

Wang’s framing was careful: he specified that the agents outperformed engineers “on specific tasks” with “the right evaluation system.” On well-defined, measurable tasks with clear success criteria, the economics are starting to look one-sided.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.