cs.CYMar 30, 2026

Synergy: A Next-Generation General-Purpose Agent for Open Agentic Web

Authors: Xiaohang Nie, Zihan Guo, Kezhuo Yang, Zhichong Zheng, Bochen Ge, Shuai Pan, Zeyi Chen, Youling Xiang, 

Organizations: Shanghai Innovation Institute · Harbin Institute of Technology · Holos Engineering · Sun Yat-sen University · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

AI agents are rapidly expanding in both capability and population: they now write code, operate computers across platforms, manage cloud infrastructure, and make purchasing decisions, while open-source frameworks such as OpenClaw are putting personal agents in the hands of millions and embodied agents are spreading across smartphones, vehicles, and robots. As the internet prepares to host billions of such entities, it is shifting toward what we call Open Agentic Web, a decentralized digital ecosystem in which agents from different users, organizations, and runtimes can discover one another, negotiate task boundaries, and delegate work across open technical and social surfaces at scale. Yet most of today's agents remain isolated tools or closed-ecosystem orchestrators rather than socially integrated participants in open networks. We argue that the next generation of agents must become Agentic Citizens, defined by three requirements: Agentic-Web-Native Collaboration, participation in open collaboration networks rather than only closed internal orchestration; Agent Identity and Personhood, continuity as a social entity rather than a resettable function call; and Lifelong Evolution, improvement across task performance, communication, and collaboration over time. We present Synergy, a general-purpose agent architecture and runtime harness for persistent, collaborative, and evolving agents on Open Agentic Web, grounding collaboration in session-native orchestration, repository-backed workspaces, and social communication; identity in typed memory, notes, agenda, skills, and persistent social relationships; and evolution in an experience-centered learning mechanism that proactively recalls rewarded trajectories at inference time.

Figures & tables

Figure 1: Overall architecture of Synergy.
Figure 2: Collaboration and execution lifecycle in Synergy. A complex task begins in a primary session, branches through Cortex-managed child sessions, moves across mailbox-mediated asynchronous delivery and repository-backed collaborative surfaces, and returns to the originating session as traceable outputs. The figure emphasizes that Synergy’s collaboration model is not only message passing, but bounded execution that can branch, delegate, re-incorporate results, and extend into shared workspaces and remote environments without losing accountability.
Figure 3: Experience learning loop in Synergy. Past experiences are actively retrieved and injected into the current task context, after which the resulting trajectory is evaluated using either explicit benchmark feedback or dialogue-derived reward from subsequent interaction. The resulting multi-dimensional reward is then used to update the reused experiences through delayed credit assignment, so that future recall becomes increasingly value-aware and the accumulated experience store becomes a reusable, partially transferable capability asset.
Figure 4: Capability growth under experience accumulation. Panels (a) and (b) show full performance trajectories on SWE-bench Verified and OpenRCA, making visible both the steady upward movement over epochs and the concentration of gains in the early stages of accumulation. Panel (c) summarizes final gains over the starting point of each accumulated-experience run, highlighting that the resulting improvements are substantial in both absolute and relative terms.
BenchmarkModelEpochsStartBestGain (pp)Relative gainGain by epoch 5Avg patch rate
SWE-bench VerifiedQwen 3.5 397B A17B1563.082.6++ 19.6++ 31.1%71.4%95.0%
SWE-bench VerifiedNex 1.11260.883.0++ 22.2++ 36.5%72.1%95.9%
OpenRCAQwen 3.5 397B A17B811.9429.6++ 17.7++ 148.1%72.7%—
Table 1: Summary of capability growth under experience accumulation. Start and Best denote the first and best observed scores within the accumulated-experience runs. Relative gain reports improvement relative to the starting point, while Gain by epoch 5 reports the fraction of total improvement already realized by the fifth epoch. For SWE-bench Verified, Avg patch rate indicates that the growth process remains operationally stable while performance rises.
Figure 5: Immediate capability gains from transferred experience on the OneMillion Benchmark. Panel (a) compares four overall metrics between the no-experience baseline and the experience-injected condition, with absolute deltas annotated above each pair. Panel (b) shows that the improvement is broad rather than domain-specific: every domain in OneMillion benefits, with gains ranging from +22.1 pp (law) to +32.7 pp (healthcare).
More in this topic

Web Agents

Web agents use language and vision models to navigate websites, gather information, and complete tasks through interactive interfaces.

See all 174 papers

Explore similar work

CardsList

  1. Position: Collaborative Agentic AI Needs Interoperability Across Ecosystems

    May 25, 2025Rishi Sharma, Martijn de Vos, Pradyumna Chari +2Agent Communication ProtocolsMulti-Agent Collaboration

    1EPFL, Lausanne, Switzerland · 2Massachusetts Institute of Technology, Cambridge, MA, USA.

  2. MolmoWeb: Open Visual Web Agent and Open Data for the Open Web

    Apr 9, 2026Tanmay Gupta, Piper Wolters, Zixian Ma +13Web AgentsLLM Agent Training

    Allen Institute for AI · University of Washington · UNC-Chapel Hill

  3. Structured Distillation of Web Agent Capabilities Enables Generalization

    Apr 9, 2026Xing Han Lù, Siva ReddyLLM Agent TrainingWeb Agents

    Mila – Quebec AI Institute McGill University