{"slug": "for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd", "title": "For coding agents, real-time collaboration beats the \"wisdom of the crowd\"", "summary": "A new experiment by Antigravity using Gemini 3.5 Flash on 30 Project Euler problems found that five AI coding agents collaborating in real time achieved 87% accuracy, outperforming five solo agents at 72% and their majority vote at 80%. The collaborating agents also solved problems faster, with average run time dropping from 14 to 10 minutes and median worst-of-5 time from 31 to 11 minutes.", "body_md": "**tl;dr:** I ran AI coding agents on 30 [Project Euler](https://projecteuler.net/) problems in two ways: five\nagents working separately, and five agents collaborating in real time. For\nboth, I also computed the majority vote across the five answers at the end.\nThe collaborating agents beat both the separate agents and their majority\nvote. Average accuracy without the vote went from 72% to 87%.\n\nI had AI coding agents, specifically Gemini 3.5 Flash running through\nAntigravity, tackle 30 math and programming problems from Project Euler: all problems\npublished this year except one ([983](https://projecteuler.net/problem=983)),\nwhose answer I could not verify on the website.\n\nFor each problem, five agents ran in\n[five concurrent containers](https://github.com/ykdojo/antigravity-cloud-run).\nI ran them in two ways:\n\nIn both cases I also took the majority vote across the five answers at the end. So in total there are four methods to compare: solo individuals, solo with vote, collaborative individuals, and collaborative with vote.\n\nEach agent had 30 minutes of working time, plus up to 10 extra minutes to wrap up and submit a final answer, so a single run can take up to about 40 minutes.\n\nIndividually, solo agents got 108 of 150 runs correct (72%). With real-time collaboration, the same setup got 130 of 150 (87%). The majority vote went from 24 of 30 problems (80%) to 27 of 30 (90%). As you can see, both real-time collaboration and the “wisdom of the crowd”, aka majority voting, are effective, but real-time collaboration had a stronger effect on the overall accuracy.\n\n| solo | collaborative | |\n|---|---|---|\n| average run time | 14m | 10m |\n| median worst-of-5 time | 31m | 11m |\n\nThe worst-of-5 time is relevant because if you wait for all five answers, that is when you get the final answer through the vote. The average time is relevant in a different way: it is a proxy for measuring the tokens and compute you use. Antigravity unfortunately does not give you exact token counts, but tokens and compute scale with the time it takes to solve a problem.\n\nMore collaborative runs end inside the first five minutes (97 vs 77), and fewer of them run past the 30 minute mark (23 vs 38).\n\nOn worst-of-5 time, solo sets pile up at the 30 to 40 minute end because many problems left at least one agent running until the cap. Collaborative sets most often finish within the first five minutes because once one agent solves the problem and shares its findings, the rest can finish quickly too, though the harder ones still run long.\n\nThree individual problems are worth calling out:\n\nThese agents are not deterministic. If you try to reproduce the results you might get slightly different numbers, but in principle you should be able to get similar results.\n\nPer Project Euler’s rules, no numeric answers are published here or in the repo. Runs were scored against privately held answers that were verified on the site itself.\n\nThe insights here were drawn from experiments with only 5 agents, so there are still many unknowns.\n\nA lot of open questions, but I believe this is a decent start.\n\nOne of the closest things currently in production is xAI’s Grok Heavy line: Grok 4 Heavy ran multiple agents in parallel that compare notes like a study group, and the newer versions make a multi-agent setup the default for complex queries.\n\nThe exact prompts, the Dockerfile, and the full per-run outcome data (correct\nor not, and time taken, for all 300 runs) are in the\n[appendix](/antigravity-cloud-run/experiments/collaboration-vs-wisdom-of-the-crowd/appendix.html).", "url": "https://wpnews.pro/news/for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd", "canonical_source": "https://ykdojo.github.io/antigravity-cloud-run/experiments/collaboration-vs-wisdom-of-the-crowd/", "published_at": "2026-07-25 23:07:49+00:00", "updated_at": "2026-07-25 23:22:12.437770+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "large-language-models"], "entities": ["Antigravity", "Gemini 3.5 Flash", "Project Euler", "xAI", "Grok 4 Heavy"], "alternates": {"html": "https://wpnews.pro/news/for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd", "markdown": "https://wpnews.pro/news/for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd.md", "text": "https://wpnews.pro/news/for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd.txt", "jsonld": "https://wpnews.pro/news/for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd.jsonld"}}