A swarm rider is a person who can consistently command large swarms of AI agents to productive ends. The term was coined by Matt Huang, co-founder of Paradigm, on Nov 10, 2026.
21 lit particles: the riders on the register so far
Where the term comes from
“There seems to be a new class of people — call them ‘swarm riders’ — who are capable of consistently commanding large swarms of agents to productive ends.”
Huang describes swarm riding as a new managerial skill that is closer to machine learning than to software engineering: defining reward functions and evals, turning non-verifiable problems into verifiable ones, and running experiments on black-box systems. Source: the essay.
How many tokens do swarm riders use?
Based on the swarm riders he knows, Huang puts their productive token spend at 1,000–10,000x that of a median software engineer and 100x that of a top-5% Claude Code user (essay). Riders who stated their own scale include Jacob Everly (8–16 agents · ~600M tokens/day (aiming for 1B)), Jacob Sansbury (6.5T Claude Opus tokens in 30 days) and Noah Zixin Zhang (500M–1B tokens/day (later 1–2B)). These numbers are self-reported.
What is a swarm-native company?
In Huang’s essay, a swarm-native company is founded by and populated by swarm riders and organized around swarms in a way conventional companies will struggle to copy, because human coordination is “sand in the gears”. He names Tempo, Spawn and Cathedral. See all swarm-native companies.
Is a fleet of agents the same as a swarm?
Not everyone thinks so. One reply to the essay by @pegzeus argued that most swarms today are really fleets run by one owner, while a true swarm is an open market anyone can join. Miguel Ríos Berríos makes a related point: a swarm is coordinated work, not a batch job over a list.
Do agent swarms always beat a single agent?
No. Anthropic argues that one agent with skills often beats a multi-agent setup on quality, cost and latency, and ships the single-agent-loop-kit to test it. SwarmChat’s experiments (CMU and Fleet) found swarms got there faster but didn’t raise the ceiling: Pranjal Aggarwal reports 16-agent swarms matched a single agent’s score 6x faster.
Examples from the register
Practitioners who have publicly described their own swarms:
Jacob EverlyProduct, Boundless8–16 agents · ~600M tokens/day (aiming for 1B)
Jacob SansburyFounder, Spawn6.5T Claude Opus tokens in 30 days