Gormes

Academic Research Survey

Survey of academic papers and technical research on building better agentic operative systems.

Agentic Operating Systems — Academic Research Survey

This directory contains a comprehensive survey of academic papers, technical research, and open-source implementations relevant to building better agentic operative systems like OpenCode, Hermes Agent, and Claude Code.

Research covers English and Chinese (中文) academic sources across 2022–2026, with findings organized by domain.

Contents

FileTopicCoverage
Foundational ArchitecturesReAct, Reflexion, ToT, Voyager, Generative AgentsCore agent loop patterns
Agentic OS DesignAIOS, ACOS, OpenCode, Hermes AgentOS-level agent architecture
Agent Memory SystemsAgeMem, Synapse, A-Mem, H-MEM, MemGenEpisodic, semantic, hierarchical memory
Multi-Agent & ToolsMetaGPT, CAMEL, MCP, ToolACEOrchestration & tool ecosystems
Safety & DeploymentMOSAIC, Agent-C, OpenSandbox, production patternsGuardrails & production reliability
Chinese Research (中文综述)智能体操作系统、多智能体协作、工具学习Complete Chinese-language survey
Reading ListCurated priority reading orderWhere to start

Key Themes

ThemeCore Insight
Agent LoopThought→Action→Observation cycle (ReAct, 2022) remains foundational; variants add reflection, tree search, and hierarchical decomposition
OS DesignAIOS (2024) establishes “LLM as OS kernel” paradigm with scheduler, context manager, tool manager, and access control
MemoryShift from passive vector stores to agentic memory with autonomous store/retrieve/forget decisions (AgeMem, 2026)
Multi-AgentMetaGPT (ICLR 2024) shows SOP-based collaboration; MACNet (ICLR 2025) verifies collaboration scaling laws
Tool SystemsMCP standardizes agent-tool interface; SWE-Agent’s ACI concept shows interface design matters as much as model quality
SafetyMOSAIC (plan→check→act loop), Agent-C (formal temporal guarantees), IntentGuard (training-free runtime gates)
DeploymentEvent-driven pub/sub beats polling; checkpointing enables pause/resume; multi-channel gateways with shared kernel

Research Methodology

Sources surveyed:

  • English: ArXiv, Google Scholar, NeurIPS, ICLR, ACL, EMNLP, ICML proceedings (2022–2026)
  • Chinese: 知网 (CNKI), 万方, ArXiv Chinese teams, CCF-recommended venues
  • Open-source: GitHub repositories, documentation, architecture RFCs

Contributing

Gormes is a Go-native Hermes-compatible agent runtime. This research informs its architecture decisions. If you find papers or systems not yet covered, open an issue or PR.