Swarm Riding: The Skill of Steering AI Agent Swarms

Swarm riding is the newest job skill nobody has on their CV yet. Matt Huang, co-founder of Paradigm, coined the term on 10 November 2026 in a post called Swarm Riders and Swarm-Native Companies. A swarm rider, in his words, is someone "capable of consistently commanding large swarms of agents to productive ends." Ten days later, Paradigm's CTO Georgios Konstantopoulos published a field guide, Effective Swarm Riding for Engineers. So the term got a definition and a manual in under two weeks. Most JavaScript frameworks wait longer for docs.
What swarm riding means
Strip off the cool name and it is this. You stop pairing with one AI agent. You run many at once and judge them on results. Huang calls it a managerial skill, closer to running a big organisation than writing code. He also says it looks more like machine learning than software engineering: you define reward functions and evals, turn fuzzy problems into ones a test can check, and run experiments.
His numbers are wild. Based on the riders he knows, their productive token spend is 1000x to 10000x that of a median software engineer, and 100x that of a top 5% Claude Code user. One reply from Will Papper said it reminded him of sandworm riding in Dune. Fair. Both involve holding on to something huge and hoping it goes your way.
A quick example
Say your API is slow. The old way: you profile, find the hot path, fix it, open a PR. The swarm way, per Konstantopoulos, is different. You do not say "Optimize X". You say "Make this 10x faster", point cloud agents at a codebase that is set up for them, and get out of the way. One prompt on his list is literally "I am stepping away for 8hrs, believe in yourself".
Your real job is the scoreboard. A benchmark the agents must beat. Tests that catch them when they cut corners.
Where I agree, and where I flinch
The core idea is right. The hard part moves from typing code to deciding what "good" means and measuring it. As a tester, I like that. Checks stop being the chore at the end and become the main work.
But parts of the guide made me put my coffee down. Konstantopoulos says to remove comments, delete docs/ folders, remove agent-written unit tests, and (he flags this one as controversial) push to master and drop CI, because "the agent thread is the new pull request & CI." To be fair, he says to add deterministic simulation tests instead, and he closes by saying effective swarm riding "ultimately reduces to code security." I agree with the ending. I am not sold on the middle. Throwing out your tests because the agents got faster feels like firing the auditors because the accountants got faster.
Then there is cost. Burning 10000x the tokens of a median engineer only pays off if the output is real. The independent tracker swarmriders.ai lists self-reported setups, including one founder who claims 6.5 trillion Claude Opus tokens in 30 days. The same site cites SwarmChat experiments where 16-agent swarms matched a single agent's score 6x faster, but did not raise the ceiling. Faster, yes. Smarter, not yet.
Huang also coined a second term in the same post: swarm-native companies. These are firms built around swarms from day one, where human coordination is "sand in the gears." He names Tempo, Spawn and Cathedral.
So here is my question for you. If your test suite is the only thing standing between a swarm and production, who writes the tests, and who checks them?