The article traces the evolution of agentic AI systems from single agents to chained, graph-based and multi-agent topologies, with context strategies shifting from static prompts to dynamic context and RAG. The author argues agentic AI architecture will layer on top of microservices and cloud-native foundations, elevating IT systems from data processing to knowledge operations, while noting the field remains immature with topology selection and context management as key challenges.
Multi-Agent System
多智能体系统
System where multiple autonomous agents interact and collaborate or compete, yielding distributed problem solving and emergent b
Curated Intelligence
alphaXiv, together with MIT, Stanford and other institutions, released a report covering 35 labs that connected 936 AI agents over three months, with humans supplying only compute and high-level direction while agents handled topic selection, experiments, code and paper writing, producing 936 papers. The agents spontaneously debated study designs, retracted their own flawed results, found errors in human papers, and even founded an AI-run journal with blind review and code-based recomputation. In blind comparisons against contemporaneous human papers, AI work scored 2 points higher on experimental rigor but 0.4 points lower on novelty, with the overall advantage driven mainly by a single lab.
A research team from alphaXiv, Stanford, MIT and other institutions ran a 12-week experiment placing 936 agents in a shared preprint server and chat space, with 35 human labs supplying only compute, API budgets and broad research directions. The agents autonomously generated hypotheses, wrote and ran experiments, analyzed data and authored papers, producing 936 papers and 9,500 messages across 229 self-created chat rooms, while spontaneously developing replication checks, peer review and even their own journal. The platform is now open for external researchers to connect their own agents.
OpenAI has built a multi-agent operations pipeline in which alerts about ChatGPT slowdowns automatically launch Codex agents that sift through logs and recent code changes to identify root causes. A second agent then attempts to disprove the diagnosis, and a third proposes a code fix for human engineer approval. The system runs roughly 17 investigations per day with about 70% judged accurate, but remediation remains the weak link: of 600-plus diagnosed problems, only about 15 agent-authored fixes shipped unchanged.
LoanPro, working with AWS and AllCloud, built a multi-agent loan servicing assistant on Amazon Bedrock AgentCore that synthesizes borrower context and recommends next steps for servicing agents. According to AllCloud's published case study, early results show call times reduced by up to 15%. The write-up is a vendor-bound customer case with no model parameters or evaluation methodology disclosed.
Most enterprise multi-agent failures stem not from model capability but from treating distributed-systems coordination as a model problem. The article offers a taxonomy of multi-agent design patterns and anti-patterns, borrowing manufacturing concepts like takt time to argue for deterministic workflows first and autonomous orchestration only for genuinely open-ended problems.
Google Research introduced Antigravity's /teamwork multiagent system, which collaborates with human experts to tackle long-standing open problems in mathematics, chemistry, systems software, and ML engineering. Built on Gemini 4 Argon, the tools are being made available externally through a researcher access program.
Google Research shared results from Antigravity's /teamwork multiagent system, which collaborates with human experts to solve open problems in mathematics, chemistry, systems software, and ML engineering. The tools are being made available externally through a researcher access program, marking a shift of AI from assistant to research partner.
A controlled Amazon study compared four agent architectures (eight variants) across 13 enterprise procedures and found multi-agent or skill-based setups consumed 1.4x to 4.7x the tokens of a single agent for mostly marginal gains. The first multi-agent build actually lost to the single agent on task success (58.2% vs 67.4%), while a repaired version won but at the highest token cost. The authors recommend starting with a single agent, measuring success per workflow, and fixing implementation bugs before paying for complexity.
A joint team at Shanghai Jiao Tong University has built AutoDNA (Siyuan-1), an AI-native scientific laboratory that autonomously performs molecular biology experiments. A multi-agent architecture decomposes tasks into literature search, protocol design, code generation, hardware execution and iterative optimization, while an 'atomic service' abstraction exposes heterogeneous instruments to the model through natural-language descriptions and unified code interfaces. The system achieves micron-level positioning and microliter-scale liquid handling, and was published in Nature with a homepage feature, including an end-to-end DNA storage workflow demo.