# AI Agent Org Forms Underperform Human-Like Hierarchy | Sekoul Krastev posted on the topic **By:** Sekoul Krastev **Published:** 2026-07-17T12:09:48.500Z **Source:** [LinkedIn](https://www.linkedin.com/posts/sekoul_the-more-systems-of-ai-agent-are-organized-share-7483855471649660928-NAav) --- AI Agent Org Forms Underperform Human-Like Hierarchy This title was summarized by AI from the post below. Sekoul Krastev 3d Report this post The more systems of AI agent are organized to resemble human org charts, the worse they seem to perform. A new preprint by Canhui LIU (UCL) ran 8,000 synthetic knowledge-work tasks across seven different agent organization forms, from pipelines and hierarchies to shared-memory blackboards and adaptive meta-organizations. The results vary quite a lot. • Adaptive meta-organization: +11.43 efficiency points, +23.24 percentage points success probability vs single expert • Blackboard (shared memory): +7.60 efficiency, +19.48 pp success • Hierarchy manager: -7.92 efficiency • Committee debate: -12.69 efficiency Committee debate performed worst of all. The agents produced "pseudo-diversity," basically just paraphrasing each other instead of generating independent evidence. The more the structure resembled the typical human team deliberation process, the worse it got. Agent-native forms were 395.26% more efficient than human-imitation forms overall. This isn't exactly a critique of humans orgs (at least in my mind it's not). There are many reasons why orgs are shaped the way they are, and short-term efficiency is hardly the end-all-be-all. That said, as using single agents evolves into creating complex systems of agents, this kind of research opens a lot of questions about how those systems should be organized. 112 9 Comments Like Comment Sekoul Krastev 3d Report this comment Read full preprint here: https://arxiv.org/abs/2606.30986 Like Reply 4 Reactions 5 Reactions Wojciech Zygmunt Kaleta, 🎓 PhD 3d Report this comment Sekoul Krastev, interesting result. I wonder whether the comparison changes once the task requires not only finding the best answer, but also establishing who has the continuing authority to act, reopen, or intervene as conditions change. Human organisations are not built only to optimise throughput. They also distribute accountability, preserve intervention rights, and keep decisions legitimate over time. The real question may not be which architecture is more efficient, but which one continues to make the right intervention possible after the environment changes. Like Reply 1 Reaction 2 Reactions Labadé O. 3d Report this comment The deeper lesson may be that innovation is often constrained by inherited organizational metaphors. When we make a new capability imitate an old structure, we preserve the structure’s weaknesses along with its familiarity; the real advantage appears when coordination is designed around the capability itself. Fascinating research. Like Reply 1 Reaction Olli Salo 3d Report this comment Super interesting! How far is Adaptive meta-orchestration from the basic behaviour of basic Claude Code when you allow it to spin up sub-agents as needed? Like Reply 1 Reaction 2 Reactions Johan Heinen 3d Report this comment I had to make a sidestep to AI to fully onderstand the org. desciptions:-). Interesting. Like Reply 1 Reaction Tristan Ingold 3d Report this comment Time to research more about adaptive meta-organizations 🤔 Like Reply 1 Reaction See more comments To view or add a comment, sign in More Relevant Posts Michael Jastram 5d Edited Report this post AI can accelerate writing, analysis, reviews and automation. But faster tasks do not automatically produce faster products. On September 9, I will speak with Marc Osofsky, CEO of Jama Software, about what turns local AI productivity into measurable Product Velocity. We will discuss: • Why task-level gains often fail to improve end-to-end flow • How engineering data and traceability affect decision speed • What spec-driven development changes across the lifecycle • How platforms can connect faster work to faster outcomes • Where AI becomes another local optimization instead of a system-level improvement This is an editorially independent SE-Trends webinar, organized and moderated by me. It is a substantive discussion, not a commercial product presentation. Wednesday, September 9 17:00 CEST | 11:00 EDT | 8:00 PDT Participation is free. Register here: https://lnkd.in/dE8afYyx #ProductVelocity #Engineering #AI #SystemsEngineering #ProductDevelopment #JamaSoftware 32 Like Comment To view or add a comment, sign in Simon Bennett 3d Report this post Sharing to reach out to my semiconductor industry colleagues and friends. I firmly believe that AI combined with systems engineering is essential to helping us achieve that $1T 2030 target set for all of us. Give this discussion a listen. Michael Jastram Product Velocity for complex physical product development | Reduce time-to-market and decision latency | Trainings, workshops & 30 day interventions 5d Edited AI can accelerate writing, analysis, reviews and automation. But faster tasks do not automatically produce faster products. On September 9, I will speak with Marc Osofsky, CEO of Jama Software, about what turns local AI productivity into measurable Product Velocity. We will discuss: • Why task-level gains often fail to improve end-to-end flow • How engineering data and traceability affect decision speed • What spec-driven development changes across the lifecycle • How platforms can connect faster work to faster outcomes • Where AI becomes another local optimization instead of a system-level improvement This is an editorially independent SE-Trends webinar, organized and moderated by me. It is a substantive discussion, not a commercial product presentation. Wednesday, September 9 17:00 CEST | 11:00 EDT | 8:00 PDT Participation is free. Register here: https://lnkd.in/dE8afYyx #ProductVelocity #Engineering #AI #SystemsEngineering #ProductDevelopment #JamaSoftware 5 Like Comment To view or add a comment, sign in Civic Roundtable 4,313 followers 3w Report this post What can AI actually do for the public servants serving our communities today? Fast Company just published a piece by our cofounder & CEO, Madeleine Smith, with one clear answer: it can help agency leaders 𝘀𝗲𝗲 𝗮𝗿𝗼𝘂𝗻𝗱 𝗰𝗼𝗿𝗻𝗲𝗿𝘀. The signal leaders need is already there, in the qualitative data most agencies lack ways to stitch together: • the questions frontline public servants are asking each other, and • the resources they're suddenly requesting. In other words, dashboards tell you what already happened — but the questions moving through the field tell you what's coming. AI that reads those signals in real time can surface leading indicators of a problem before it becomes a full-blown crisis. Making this shift can free up agencies from being reactive, to working more proactively. Read the full piece here: https://lnkd.in/g_-U6GGq 4 Like Comment To view or add a comment, sign in Uniphore 86,890 followers 1w Edited Report this post Every enterprise now has access to the most powerful Large Language Models (LLMs), but organizations really achieve competitive advantage when the models know specific context about your business. We’re seeing organizations pull ahead with connected proprietary data, institutional knowledge and context. The results include: → Better decisions → Stronger automation → Outcomes that other competitors cannot replicate Uniphore CEO Umesh Sachdev discusses why enterprise AI success is highly contingent on data, context and agent training – cornerstones at the center of enterprise AI strategy. The real challenge enterprises face is turning decades of company knowledge into operational intelligence. The critical question: how do you transform proprietary data into a strategic asset without losing control of it? Watch the full video here: https://lnkd.in/gYjKmbQ3 …more Enterprise AI wins with data, context and control | Uniphore 87 1 Comment Like Comment To view or add a comment, sign in Zach Breimayer 1w Report this post I'd add that SLMs are becoming a key part of this conversation. Beyond training on proprietary structured and unstructured data, purpose built Small Language Models can often deliver greater accuracy for specific business tasks, significantly lower latency, and a much lower cost profile than relying exclusively on generic frontier models. The real opportunity is matching the right model to the right workload while keeping enterprise knowledge secure, governed, and actionable. Uniphore 86,890 followers 1w Edited Every enterprise now has access to the most powerful Large Language Models (LLMs), but organizations really achieve competitive advantage when the models know specific context about your business. We’re seeing organizations pull ahead with connected proprietary data, institutional knowledge and context. The results include: → Better decisions → Stronger automation → Outcomes that other competitors cannot replicate Uniphore CEO Umesh Sachdev discusses why enterprise AI success is highly contingent on data, context and agent training – cornerstones at the center of enterprise AI strategy. The real challenge enterprises face is turning decades of company knowledge into operational intelligence. The critical question: how do you transform proprietary data into a strategic asset without losing control of it? Watch the full video here: https://lnkd.in/gYjKmbQ3 Enterprise AI wins with data, context and control | Uniphore Like Comment To view or add a comment, sign in Sarankumar R 1w Report this post Most Claude users will waste Fable -5 before it leaves. Not because the model isn't powerful. Because they're using it the wrong way. If you're using Fable -5, don't spend your remaining access on simple questions. Use it where it creates the most value. Here's how I'd use it. 🧠 1. Use Fable -5 for strategy. Ask it to solve your hardest problems. Business decisions. Product ideas. Research. System design. Not quick questions. ❓ 2. Let it interview you first. Instead of giving one prompt, try this: "I need help with [goal]. Before you answer, ask me everything you need to know." Better context. Better results. 💬 3. Start a new chat for every new task. Don't mix unrelated work. Clean context leads to better responses. 📁 4. Save your best workflows. When you get an excellent result, turn it into reusable instructions or a Claude Skill. That way, you won't need to rebuild the same workflow again. The biggest lesson? The best AI users don't ask more questions. They design better workflows. Models will change. The way you work with them is what gives you a long-term advantage. What's your favorite Claude feature right now? 💬 Share it in the comments. ❤️ Like this post if you found it useful. ♻️ Repost to help more people get better results with Claude. 🚀 Follow Sarankumar R for daily AI workflows, Context Engineering, and practical AI insights. 1 Like Comment To view or add a comment, sign in Marco Patrone 3w Report this post I have published a new working paper on Zenodo: Representation Governance Institutional Control, Protocol Coordination, and Allocative Authority in AI-Mediated Markets DOI: 10.5281/zenodo.20930753 This paper introduces Representation Governance as a distinct governance layer for AI-mediated markets. The central argument is that when AI systems increasingly mediate discovery, comparison, ranking, recommendation, and decision-making, market access is no longer determined only by websites, advertising, platforms, or human attention. It is increasingly determined by representation. If an entity is not represented in a machine-readable, verifiable, and admissible form, it may simply not enter the field of consideration. This creates a new form of authority: Not only institutional authority. Not only algorithmic authority. But representation authority. The authority to define what exists computationally before it becomes visible economically. The paper develops this framework through three concepts: Institutional Authority — who defines the formal rules. Representation Authority — who defines admissible machine-readable representations. Allocative Authority — who shapes what AI systems can surface, compare, recommend, or exclude. This is part of my broader work on the Representation Economy: the transition from web visibility to computational admissibility. In the web economy, being online was enough to exist. In AI-mediated markets, being online may no longer be sufficient. You need to be representable. Zenodo DOI: 10.5281/zenodo.20930753 Project page: https://lnkd.in/eQEysJ88 #RepresentationEconomy #AIGovernance #ArtificialIntelligence #DigitalMarkets #AIInfrastructure #MarketDesign #ComputationalEconomy #Research 2 Like Comment To view or add a comment, sign in ZBrain 393 followers 4w Report this post 💠 Enterprise teams generate and rely on vast amounts of knowledge, documents, workflows, system data, and historical insights, but connecting this knowledge to AI workflows and decision-making is often fragmented. Explore how ZBrain Builder addresses this: https://lnkd.in/gzey2zrh 💠 ZBrain Knowledge Base centralizes and organizes enterprise knowledge, enabling teams to discover, access, and operationalize insights for AI initiatives. 💠 By integrating structured and unstructured data, the Knowledge Base ensures AI agents and workflows work with contextual, up-to-date, and accurate information. Key capabilities include: ✅ Centralize documents, workflows, and system knowledge in one repository ✅ Provide AI agents with contextual access to enterprise information ✅ Support human-in-the-loop decision-making with verified knowledge ✅ Maintain governance, versioning, and auditability of knowledge assets 💠 ZBrain Knowledge Base bridges the gap between enterprise knowledge and AI execution, helping organizations unlock actionable insights, improve decision quality, and scale AI-driven workflows confidently. 👉 Explore ZBrain Knowledge Base: https://lnkd.in/gjzePHVV #ZBrainKnowledgeBase #EnterpriseAI #AgenticAI #KnowledgeManagement #OperationalExcellence #WorkflowAutomation #GovernedAI #ContextualAI …more 8 Like Comment To view or add a comment, sign in Michael Jastram 5d Edited Report this post Interview with Michael Jastram and Marc Osofsky AI can accelerate writing, analysis, reviews and automation. But faster tasks do not automatically produce faster products. On September 9, I will speak with Marc Osofsky, CEO of Jama Software, about what turns local AI productivity into measurable Product Velocity. We will discuss: • Why task-level gains often fail to improve end-to-end flow • How engineering data and traceability affect decision speed • What spec-driven development changes across the lifecycle • How platforms can connect faster work to faster outcomes • Where AI becomes another local optimization instead of a system-level improvement This is an editorially independent SE-Trends webinar, organized and moderated by me. It is a substantive discussion, not a commercial product presentation. Wednesday, September 9 17:00 CEST | 11:00 EDT | 8:00 PDT Participation is free. Register here: https://lnkd.in/dE8afYyx #ProductVelocity #Engineering #AI #SystemsEngineering #ProductDevelopment #JamaSoftware #ProductVelocity 7 Like Comment To view or add a comment, sign in 20,802 followers 891 Posts View Profile Follow Explore content categories Career Productivity Finance Soft Skills & Emotional Intelligence Project Management Education Technology Leadership Ecommerce User Experience