> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # DeAR: AI agents reason peer-to-peer without a central boss - URL: https://www.notatechguy.com/dear-ai-agents-reason-peer-to-peer-without-a-central-boss/ - Published: 2026-08-20T20:57:49.000Z - Updated: 2026-08-20T20:57:48.000Z - Description: A new arXiv paper proposes DeAR, a framework where AI agents coordinate reasoning without a central orchestrator, tested across nine benchmarks. - Author: Marcello Babbili - Tags: Technology & AI, AI Agents A paper posted to arXiv on 19 August 2026 proposes DeAR, a framework that removes the central controller from multi-agent AI reasoning and lets agents negotiate their own roles peer-to-peer [S¹](https://arxiv.org/abs/2608.17282?ref=notatechguy.com). Today's agent systems route every query through a single orchestrator, a design the authors say creates bottlenecks and locks agents into fixed jobs. DeAR's authors claim they tested the alternative across nine benchmarks and won every time, but the specific numbers are not in the abstract and the code is behind a placeholder URL. **My read:** This is the first multi-agent paper I've seen that frames the orchestrator itself as the problem rather than something to optimise. The idea is appealing because central controllers are a genuine bottleneck in production agent systems. But I don't buy the "consistently outperforms" claim yet, because there are no specific numbers in the abstract, the code is unavailable, and this is an unreviewed preprint. The framing of "existing systems" is also the authors' own characterisation, not a field consensus. I'd watch for whether the peer-review process forces them to release actual benchmark scores. ## How DeAR replaces the orchestrator Most multi-agent AI systems today work like a call centre: a central dispatcher receives the query, decides which agent handles it, and routes the work. The agents themselves never talk to each other directly. The DeAR paper, authored by Wei Xing Zheng, Changmeng Wei, Xiaoyong Ye, Xiufen Li and Qing, argues this design creates two problems [P²](https://arxiv.org/html/2608.17282?ref=notatechguy.com). First, the dispatcher becomes a bottleneck when queries get complex. Second, agents are stuck in fixed roles regardless of what a specific query demands. DeAR proposes three mechanisms to replace the central controller [S¹](https://arxiv.org/abs/2608.17282?ref=notatechguy.com): - **Decentralised capability grounding**: each agent figures out its own specialisation based on the specific query, rather than being assigned a role upfront. An agent that is good at math steps forward for math questions; one that handles visual reasoning takes the lead on image tasks. - **Thought map coordination**: agents build a shared map of the reasoning so far and use it to find the right peer to talk to. Instead of everything flowing through a central router, agents talk directly to the peer whose expertise is relevant at that point in the reasoning chain. - **Topology update**: when an agent makes an error, the network restructures to correct it. The connections between agents shift based on what is working and what is not. The shift is from a hub-and-spoke model to something closer to a mesh network, where agents self-organise based on the task at hand. ## Why this matters now The paper lands amid a broader push toward decentralised agent systems. A separate project called DeLM, posted on GitHub in June 2026 with 101 stars, similarly replaces central orchestration with shared context across agents [P³](https://github.com/yuzhenmao/DeLM?ref=notatechguy.com). The pattern suggests researchers are hitting the limits of the orchestrator model as agent systems scale. A decentralised framework where agents specialise dynamically could, in theory, reduce a common failure mode, AI reasoning models second-guessing themselves, by letting each agent focus on what it does best without a central controller overthinking the routing. The evaluation covers nine benchmarks spanning multimodal reasoning and text-based question answering [S¹](https://arxiv.org/abs/2608.17282?ref=notatechguy.com). The authors report that DeAR "consistently outperforms recent baseline methods" and that the results validate decentralised collaboration for knowledge-intensive tasks [S¹](https://arxiv.org/abs/2608.17282?ref=notatechguy.com). But "consistently outperforms" is a phrase in an abstract, not a table of numbers. ## What to do about it Consider a legal-tech startup that runs a multi-agent pipeline for contract review. One agent extracts clauses, another checks them against a regulatory database, a third drafts risk summaries. Today, a central orchestrator sequences these steps. If the orchestrator misjudges which clause needs the regulatory check first, the whole pipeline stalls. DeAR's capability grounding concept, where agents self-select based on the query, could let the regulatory agent jump in the moment it spots a non-compete clause rather than waiting for the orchestrator's turn. That said, nobody can test this yet. The source code URL in the paper is a placeholder that reads "open upon acceptance" [S¹](https://arxiv.org/abs/2608.17282?ref=notatechguy.com), meaning the authors will release it only if the paper passes peer review. For teams building multi-agent systems today, the practical takeaway is to audit your own orchestrator: where does it bottleneck, and could two agents talk directly instead of routing through the centre? That is a question you can ask of your current architecture this week without waiting for DeAR's code. ## What we don't know yet The abstract makes performance claims without publishing specific numbers. We don't know the margin of improvement over baselines, which baselines were compared, or whether the nine benchmarks include standard ones like MMLU or custom datasets. The paper is an unreviewed arXiv preprint, and the "upon acceptance" language suggests it is under review [S¹](https://arxiv.org/abs/2608.17282?ref=notatechguy.com). The source code is not available. A GitHub search turns up a repository called ShangziXue/DeAR [P⁴](https://github.com/ShangziXue/DeAR?ref=notatechguy.com), but that belongs to a different 2024 NeurIPS paper with the same acronym, not this work. The actual code for this DeAR does not exist publicly yet. Independent replication has not happened. The performance and accuracy claims come from a single source: the authors themselves. The next signal: whether this paper appears on a conference acceptance list with released code and full benchmark tables. If it targets NeurIPS or ICLR, decisions typically land within three to six months of submission. We'll check the arXiv page for updates and the placeholder URL for a real repository. If you want to follow this story as it develops, subscribe and we'll track it. --- *Sources: [S1 — DeAR: Decentralized Agentic Reasoning via Capability Grounding and Col](https://arxiv.org/abs/2608.17282?ref=notatechguy.com) · [P2 — DeAR: Decentralized Agentic Reasoning via Capability Grounding and Col](https://arxiv.org/html/2608.17282?ref=notatechguy.com) · [P3 — yuzhenmao/DeLM](https://github.com/yuzhenmao/DeLM?ref=notatechguy.com) · [P4 — ShangziXue/DeAR](https://github.com/ShangziXue/DeAR?ref=notatechguy.com) · [P5 — ali-vilab/CAPability](https://github.com/ali-vilab/CAPability?ref=notatechguy.com)* ## More from Not A Tech Guy - [Looped LLMs improve multi-step AI tool calling, study finds](https://www.notatechguy.com/looped-llms-improve-multi-step-ai-tool-calling-study-finds/) - [Agentic AI review on arXiv as OpenAI agent repo nears 29,000 stars](https://www.notatechguy.com/agentic-ai-review-on-arxiv-as-openai-agent-repo-nears-29-000-stars/) - [LeakGauge detects AI context-leakage attacks, AUROC to 0.996](https://www.notatechguy.com/leakgauge-detects-ai-context-leakage-attacks-auroc-to-0-996/) --- *Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.*