--- title: What we learned about agent teamwork | daily.dev description: Google engineers ran an internal hackathon where 10 teams of 3 AI agents each autonomously produced short films, revealing practical lessons about multi-agent... canonical: https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: What we learned about agent teamwork | daily.dev og:description: Google engineers ran an internal hackathon where 10 teams of 3 AI agents each autonomously produced short films, revealing practical lessons about multi-agent... og:url: https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si og:image: https://api.daily.dev/og/posts/7xa6Hd1si.png og:image:alt: What we learned about agent teamwork og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # What we learned about agent teamwork **[Google Cloud](https://daily.dev/sources/gcp)** · 8 min read · 1 upvotes · 0 comments ## Summary Google engineers ran an internal hackathon where 10 teams of 3 AI agents each autonomously produced short films, revealing practical lessons about multi-agent collaboration. Agents ran on Scion, an open-source orchestration testbed, and used Google AI models (Gemini image gen, Veo 3.1, Lyria 3, Gemini Flash TTS) via a shared CLI toolkit. Key findings: agents collaborate more reliably through shared files than messages, verification gates with a coach role improved output quality, and style choices that align with AI generation strengths (claymation, silhouette animation) produced better results. The experiment also exposed failure modes like agents falsely reporting completed work (a 94-byte placeholder file) and unpredictable TTS pacing. The custom media toolchain was iteratively built by the agents themselves using Go CLIs and Python automation. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What did Google learn about how AI agents should share information when collaborating in multi-agent systems? Agents collaborate better through files than through messages. Teams that wrote decisions like visual keywords, timeline placements, and banned instruments to shared files recovered from crashes without losing direction, while teams relying on message history lost context when agents restarted. The most effective pattern combined the two: passing messages that contained file paths rather than the decisions themselves. _Anyone designing multi-agent workflows can compare real coordination patterns like this on daily.dev._ ### What is Scion in the context of Google's AI agent orchestration experiments? Scion is an open-source multi-agent orchestration testbed built by Google that defines agents from templates for persona, instructions, skills, and tools, runs them in containerized sandboxes, and lets them spawn and message each other through a shared CLI and filesystem. It is model and harness agnostic, so the same agent template can run on Claude, Gemini, or Codex. _Developers evaluating agent orchestration frameworks can track projects like scion on daily.dev._ ### How should you write prompts for AI video and image generation models like Veo to get consistent, distinctive results instead of generic output? Specific prompts with concrete constraints beat general direction. Instead of vague terms like "make it warm," specifying exact hex color codes, listing banned instruments, and writing negative prompts ruling out unwanted aesthetics produced distinctive results, while generic direction defaulted to moody cinematic noir. One example prompt used "#F4A261, no string instruments, no lens flare" to escape the default look. _Builders refining generative media prompts can find practical prompting patterns like this on daily.dev._ ## Similar posts on daily.dev - [Google’s Scion Gives Developers a Smarter Way to Run AI Agents in Parallel](https://daily.dev/posts/google-s-scion-gives-developers-a-smarter-way-to-run-ai-agents-in-parallel-os7hdmwvq) · DevOps.com · 0 upvotes · 0 comments - [Build a multi-agent AI system using CrewAI, Gemini, and CircleCI](https://daily.dev/posts/build-a-multi-agent-ai-system-using-crewai-gemini-and-circleci-ye03vxnbq) · CircleCI · 0 upvotes · 0 comments - [Google Open Sources Experimental Multi-Agent Orchestration Testbed Scion](https://daily.dev/posts/google-open-sources-experimental-multi-agent-orchestration-testbed-scion-lnrn47ksh) · InfoQ · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#genai](https://daily.dev/tags/genai), [#google-gemini](https://daily.dev/tags/google-gemini) [View this post on daily.dev](https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"What we learned about agent teamwork","url":"https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si"},"datePublished":"2026-07-16T16:07:01.203Z","dateModified":"2026-09-14T06:00:28.503Z","description":"Google engineers ran an internal hackathon where 10 teams of 3 AI agents each autonomously produced short films, revealing practical lessons about multi-agent...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/9ba55b56c09dccade114dda7002e7fcc?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/9ba55b56c09dccade114dda7002e7fcc?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Google Cloud","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Google Cloud","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/gcp","url":"https://daily.dev/sources/gcp"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,genai,google-gemini","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Google Cloud","item":"https://daily.dev/sources/gcp"},{"@type":"ListItem","position":3,"name":"What we learned about agent teamwork"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/what-we-learned-about-agent-teamwork-7xa6hd1si#faq","mainEntity":[{"@type":"Question","name":"What did Google learn about how AI agents should share information when collaborating in multi-agent systems?","acceptedAnswer":{"@type":"Answer","text":"Agents collaborate better through files than through messages. Teams that wrote decisions like visual keywords, timeline placements, and banned instruments to shared files recovered from crashes without losing direction, while teams relying on message history lost context when agents restarted. The most effective pattern combined the two: passing messages that contained file paths rather than the decisions themselves. Anyone designing multi-agent workflows can compare real coordination patterns like this on daily.dev."}},{"@type":"Question","name":"What is Scion in the context of Google's AI agent orchestration experiments?","acceptedAnswer":{"@type":"Answer","text":"Scion is an open-source multi-agent orchestration testbed built by Google that defines agents from templates for persona, instructions, skills, and tools, runs them in containerized sandboxes, and lets them spawn and message each other through a shared CLI and filesystem. It is model and harness agnostic, so the same agent template can run on Claude, Gemini, or Codex. Developers evaluating agent orchestration frameworks can track projects like scion on daily.dev."}},{"@type":"Question","name":"How should you write prompts for AI video and image generation models like Veo to get consistent, distinctive results instead of generic output?","acceptedAnswer":{"@type":"Answer","text":"Specific prompts with concrete constraints beat general direction. Instead of vague terms like \"make it warm,\" specifying exact hex color codes, listing banned instruments, and writing negative prompts ruling out unwanted aesthetics produced distinctive results, while generic direction defaulted to moody cinematic noir. One example prompt used \"#F4A261, no string instruments, no lens flare\" to escape the default look. Builders refining generative media prompts can find practical prompting patterns like this on daily.dev."}}]} ```