Agent Systems Research and Evidence
智能体系统研究与证据
Contracts — explicit interfaces and reproducible fixtures can constrain named behavior; they do not make an underlying model deterministic
Automation — this repository deploys static Pages; Jules research generation and maintainer review are separate from deployment
Repository — a bilingual research and evidence base, not an implemented autonomous-agent runtime
Four Domains → Architecture + Memory + Tools + Collaboration; mechanisms retain their claim-specific implementation and validation states
Executable Evidence → Standard-library Python schema/contract checks validate documentary structure, not agent behavior
契约——明确接口与可复现夹具可以约束具体行为，但不使底层模型成为确定性系统
自动化——本仓部署静态 Pages；Jules 研究生成与维护者审查独立于部署
仓库——双语研究与证据基础，不是已实现的自治智能体运行时
四域 → 架构＋记忆＋工具＋协作；各机制保留具体声明的实现与验证状态
可执行证据 → Python 标准库 Schema 与契约检查验证文档结构，不证明智能体行为
Five Warehouses五仓
Landscape — five repository roles, not one deployed runtime
布局——五仓职责，不代表单个已部署运行时
Four Domains四域
Research — architecture, memory, tools, collaboration
研究——架构、记忆、工具、协作
Four Phases四相
Metaphor — matter phase transitions as structural narrative
隐喻——物质相变作为结构叙事
Evidence Contract
证据合同
◉ Claim State / 声明状态
◉ Evidence Level / 证据等级
◉ Mapping State / 映射状态
◉ Implementation State / 实现状态
◉ Validation State / 验证状态
Five orthogonal axes are recorded independently, not five alternative evidence statuses. A supported paper may remain a design analogy, reference-only and not tested locally. Canonical vocabulary: FOUNDATION/EVIDENCE.md; allowed values: claim.schema.json. Missing metadata does not imply verification.
五个正交轴分别记录，不是五种互斥证据状态；论文得到支持，仍可只是设计类比、仅供参考且未在本仓测试；以 FOUNDATION/EVIDENCE.md 和 claim.schema.json 的词汇为准；缺少元数据不能视为已验证
THREE PERSPECTIVES ON THE SAME SYSTEM
三个视角说的是同一件事
Engineering Positioning
工程定位
1. Paradigm Shift: From Prompt Engineering to Harness Engineering
Agent = Model + Harness is a design perspective. Explicit interfaces, permissions, stop conditions and retained evidence can make a named system more inspectable. This repository records these requirements; it does not demonstrate that prompts, context pollution or instruction drift have been eliminated.
2. Architecture Decoupling: From Execution to Encapsulation
Skills, tool interfaces and manager/sub-agent topologies are candidate design mechanisms. Their tradeoffs depend on the task, communication contract and failure model. They are not deployed here and do not establish a cure for long-context errors or memory confusion.
3. Deterministic Governance: Confronting Native Risks
Least privilege, isolation, explicit review and fail-closed checks are requirements to evaluate in a concrete implementation. MCP alone does not create process isolation. This repository's Pages workflow is deployment, not a runtime-validation CI or a proof of convergence.
1. 范式转移：从“提示词工程”到系统驾驭 (Harness Engineering)
Agent = Model + Harness 是设计视角；明确接口、权限、停止条件与留存证据，可以使特定系统更可审查；本仓记录这些要求，不宣称已经消除提示词风险、上下文污染或指令漂移
2. 架构解耦：从执行调用到能力封装 (Skill & Sub-Agent)
Skills、工具接口和 Manager-SubAgent 拓扑是候选设计机制，取舍依赖任务、通信契约与故障模型；本仓未部署这些机制，也不宣称它们根治长上下文错误或记忆混淆
3. 确定性治理：直面原生风险 (Deterministic Governance)
最小权限、隔离、明确审查与 fail-closed 检查，是需要在具体实现中评估的要求；MCP 本身不提供进程隔离；本仓 Pages 工作流负责部署，不是运行时验证 CI，也不证明收敛
Foundational Contribution
理论新增
This page is not a telemetry board, and it is not the final mirror of the continuum. Its new role is to isolate the missing theoretical layer between model capability and system behavior: the harness as a first-class object. The model may generate; the harness defines what can be remembered, invoked, rejected, sealed, or allowed to continue.
In this repository, an Agent is treated as a constrained dynamical system rather than a chat interface. The foundation does not compete with Axiom-0 or reflective-continuum; it gives both of them the vocabulary they later formalize and observe.
这个页面不是遥测面板，也不是连续体的最终镜面这里新增的职责，是把模型能力与系统行为之间缺失的一层单独抽出来：把 Harness 当成一等研究对象模型可以生成，但 Harness 决定什么可以被记忆、调用、拒绝、封存，以及继续向下流动
在这个仓库里，Agent 不被视为聊天界面，而被视为受约束的动力系统；它不与 Axiom-0 或 reflective-continuum 竞争；它给后两者提供可被形式化、可被观察的理论词汇
Ontology本体层
Defines Agent, Harness, Boundary, Memory, Tool, Collaboration, and Entropy as separate objects
把 Agent、Harness、边界、记忆、工具、协作、熵分离成独立对象
Constraint Grammar约束语法
Turns prompts and workflows into transitions, gates, rejection rules, and audit surfaces
把提示词和流程转写成状态转移、闸门、拒绝规则与审计面
Research Bridge研究桥
Maps learning theory into memory, tool use, collaboration, and architecture-level hypotheses
把学习理论映射到记忆、工具使用、协作与架构级假设
Core Philosophy
核心哲学
We distinguish — requirements from implementation. Energy functions and boundary conditions are research models whose assumptions and local mappings must be stated
We preserve — the scope of each result. Convex optimization and Lyapunov arguments apply only to systems satisfying their formal hypotheses
We require — implementation-specific validation. A paper's bound does not prove this repository's agent convergence, safety or resistance to catastrophic forgetting; this repository contains no such agent runtime
我们区分——要求与实现；能量函数和边界条件是研究模型，必须明确其假设与本地映射
我们保留——每项结果的适用范围；凸优化和李雅普诺夫论证只适用于满足其形式假设的系统
我们要求——面向具体实现的验证；论文中的界不能证明本仓智能体收敛、安全或抵抗灾难性遗忘；本仓没有这样的智能体运行时
Research Matrix
研究矩阵
| Direction方向 |
Component组件 |
Theoretical Role理论作用 |
| SimCLR + VICReg + UnsupervisedSimCLR + VICReg + 无监督 |
Memory记忆 |
Reference mechanism for learned representations; no local memory implementation学习表征的参考机制；本仓未实现记忆运行时 |
| RL (NLP) — Reversed强化学习(NLP) — 逆向 |
Tool工具 |
Candidate tool-selection/control mapping; no local tool-control implementation工具选择与管控的候选映射；本仓未实现工具控制面 |
| Federated + Spatiotemporal联邦 + 时空 |
Collaboration协作 |
External distributed-learning results under their own assumptions; no local protocol保留原始假设的外部分布式学习结果；本仓未实现对应协议 |
| Gradient Entropy (FIM/NTK)梯度熵(FIM/NTK) |
Architecture架构 |
Research analogy for optimization behavior; not a local stability guarantee优化行为的研究类比；不是本仓稳定性保证 |
Five Repositories
五仓
Foundation
agent-foundations
Historical Cross-Repository Timeline — not current implementation evidence
跨仓历史时间线——不代表当前实现证据
Shared project stage: 2026-02-12 → 2026-08-31 · 200 elapsed days. This measures the cross-repository project stage, not the age of each repository or uninterrupted automated execution.共同项目阶段：2026-02-12 → 2026-08-31 · 200 个经过日；这是跨仓项目阶段，不代表每个仓库的建仓时长或自动化连续运行天数
02/12
Maintainer-defined homepage transition — welcome-to-github; not repository creation
维护者确认的主页转型日——welcome-to-github；不是建仓日期
02/15
NEXUS CORTEX born — 846 lines, 18 files, directed knowledge graph + OODA loop
NEXUS CORTEX诞生——846行18文件，有向知识图谱+OODA循环
03/05
Severed NetworkX — zero dependency兑现; pure SQLite graph implementation
斩断NetworkX——零依赖兑现；纯SQLite图实现
03/06
Graph-Augmented Synaptic Search operational
图增强突触检索上线
03/17
Liquid warehouse migrated — FTS5 + BM25 three-stage pipeline + CognitiveReranker
液态仓迁移——FTS5+BM25三级管线+认知重排器
03/25
V3 — Merkle Chain + Recursive CTE + PageRank
V3——Merkle链+递归CTE+PageRank
04/04
Bionic microkernel five-layer separation + zero-copy multi-core PageRank + HMAC-SHA256
仿生微核五层分离+零拷贝多核PageRank+HMAC-SHA256
04/20
Plasma warehouse independent — pure logic constraints, 10-node DAG, 17 ADRs, zero implementation code
等离子仓独立——纯逻辑约束，10节点DAG，17个ADR，零实现代码
05/19
Five-warehouse system complete — Gas + Foundation enter the continuum
五仓体系闭环——气态+学术骨架入仓
Limits and Acknowledgments
边界与致谢
-
Google Jules + Gemini — AI-assisted research and drafting; outputs require source checks and maintainer review, not presumed truth or verified hardware provenance
Google Jules + Gemini——辅助研究与起草；输出需要来源核验和维护者审查，不自动构成真理或已验证的硬件来源
-
GitHub Actions — Static Pages deployment in this repository; research-task execution and merge approval remain separate
GitHub Actions——本仓用于静态 Pages 部署；研究任务执行和合并审批是独立环节
-
Academia and research institutions — For source-specific research whose assumptions and limitations remain attached
学术界和研究机构——提供须保留各自假设和局限的研究来源
-
The open-source community and GitHub
开源社区和GitHub