Jargon:Parallel agents
3 min read
Say this at the party
“Several AI helpers working at the same time, each on its own piece of a job or each having a go at the same one. Then the best work gets picked.”
In plain English
An AI agent is an AI that doesn't just chat. It does jobs step by step, using tools like web search. Parallel just means at the same time. So parallel agents are several AI helpers working at once, instead of one AI doing everything in a row.
There are two main flavours. In the first, a lead AI splits a big job into parts and hands one part to each helper, then puts the answers together. In the second, several AIs all try the same job, and the best answer gets picked or the answers get blended into one.
Party version: instead of one AI doing ten jobs one after another, you get ten AIs doing one job each, all at the same time. It can be faster and better, but it costs more, and someone still has to check the results.
Where it came from
Nobody owns the phrase. It's two ordinary words, and computers doing several things at once is a decades-old idea. In the AI agent sense, Anthropic wrote on 13 June 2025 about starting "parallel agents that search for information simultaneously". The day before, Walden Yan of Cognition argued the opposite in a post called "Don't Build Multi-Agents". Andrew Ng gave it a push in his newsletter The Batch on 27 August 2025, calling parallel agents "an important new direction for scaling up AI". Then the coding app Cursor added a screen for running many agents at once in Cursor 2.0, on 29 October 2025.
See it in action
Anthropic asked its research system to find every board member of the tech companies in the S&P 500 (a list of big American firms). The parallel version split the job into small searches and found the right answers. A single AI, plodding through one search at a time, failed. Bear in mind these were Anthropic's own tests, not independent ones.
What people get wrong
- More agents doesn't always mean smarter. It works best when a job splits into parts that don't depend on each other, like research. Lots of jobs don't split that neatly.
- They don't work like a team in an office. Often they can't see what the others are doing. Cognition's example: one agent builds a game's background, another builds the bird, and the two don't match.
- It isn't free speed. Anthropic found a team of agents used about 15 times as many tokens (the units AI use is measured in) as a normal chat, so it's only worth it for jobs that justify the cost.
- Someone still has to check the work. Cursor says reviewing and testing what the agents make has become the new hold-up.
Where Ugenie fits
Ugenie hands your tasks out to the AIs you already use, then brings you only the three that need you next. Like parallel agents at their best, the AIs do the legwork and you pick the winner.
Get Ugenie →Go deeper
- ArticleAndrew Ng, DeepLearning.AI (The Batch), 2025Agents Running in Parallel Get There FasterAndrew Ng on why running AIs side by side is a big new direction, and why cheaper AI makes it possible.
- ArticleAnthropic, 2025How we built our multi-agent research systemHow a lead AI hands out searches to helpers at once, with the S&P 500 example and the honest costs.
- PostWalden Yan, Cognition, 2025Don't Build Multi-AgentsThe sceptic's view: why agents that can't see each other's work end up building pieces that don't fit.
- PostCursor Team (Anysphere), 2025Introducing Cursor 2.0 and ComposerA coding app that lets you run many agents at once, and admits checking their work is the new hold-up.
- PostSimon Willison, 2025Anthropic: How we built our multi-agent research systemA well-known programmer explains why he went from doubtful to convinced after reading Anthropic's write-up.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.