A2A Protocol
Agent2Agent Protocol
Open standard for cross-framework agent communication
Knowledge Atlas
Every paper, blog post, and design doc ingested, distilled, and cross-indexed into concepts, entities, and sources — an LLM-maintained knowledge graph that grows with the project.
134 Concepts 47 Entities 82 Sources
01 Exhibits — 181
ACI — Agent-Computer Interface
Agent-Computer Interface
Design interfaces for models the way HCI designs for humans
Action Metamorphic Relations
Action Metamorphic Relations
Semantic-equivalent transforms · terminal-state equality · measurable robustness
Activation Intervention
Activation Intervention
Edit mid-layer activations · prove a representation's causal role
Agent Card
Agent Capability Discovery
An agent's digital business card · /.well-known/agent-card
Agent Interoperability
Agent Interoperability
From silos to a network · cross-framework agent collaboration
Agent-Native Software
Agent-Native Software
Design software with agents as a first-class user type
Agent OS
Agent Operating System
An OS abstraction that natively understands agent semantics
Agent Reliability Evaluation
Agent Reliability Evaluation
pass@1 ≠ reliability · the four-metric framework
Agent Resource Control
Agent Resource Control
Extend cgroup to tokens · API quotas · tool permissions
Agent Sandboxing
Agent Sandboxing
Dynamic code · semantic attack surface · layered defense in depth
Agent Scheduling
Agent Scheduling
System-layer arbitration for multi-agent LLM resource sharing
Agent Skills
Agent Skills Standard
Discoverable domain-knowledge packaging · SKILL.md at the core
Agentic Systems
Agentic Systems
Workflows vs Agents · from one-shot call to autonomous decision
Three AI Summers
Three AI Summers
Kautz's historical cycle · co-evolution of neural and symbolic
AIOS
AIOS — LLM Agent Operating System
Rutgers 2024 · six kernel modules · most complete LLM-OS analogy · COLM 2025
Aitia — Cause and Explanation
Aitia (Aristotle)
Greek philosophy · cause-as-explanation · foundation of the four causes
Allen Newell
Allen Newell
1927–1992 · Turing Award 1975 · PSSH · GPS · SOAR unified cognition
Andrej Karpathy
Andrej Karpathy
ex-Tesla AI Director · ex-OpenAI · Eureka Labs · LLM-OS architecture · Software 3.0
Answer.AI
Answer.AI
Founded by Jeremy Howard · llms.txt standard · fast.ai extension · open, practical AI tools
Anthropic
Anthropic
Claude series · interpretability leader · Agent SDK · MCP · Managed Agents
Aristotle
Aristotle
384–322 BCE · Four Causes · teleological explanation · knowledge as causal knowledge · founder of Western scientific method
Artur d'Avila Garcez
Artur d'Avila Garcez
City, University of London · neurosymbolic computing · Kautz taxonomy · XAI fidelity critique · Third Wave
ASPLOS
ASPLOS
top systems conference · AgenticOS Workshop 2026 · agent infrastructure enters the systems research agenda
Augmented LLM
Augmented LLM
Retrieval + tools + memory · the building block of agentic systems
Autonomy Slider
Autonomy Slider
From Tab-complete to whole-repo agent · designing partial-autonomy products
Avatamsaka Sutra
华严经 · Avatamsaka Sutra
Dharmadhatu arising · Indra's Net · Sumeru in a mustard seed · one-is-all · forerunner of fractal philosophy
Backdoor Criterion
Backdoor Criterion
Graphical deconfounding criterion · Pearl's backdoor adjustment
Bag of Heuristics
Bag of Heuristics
Local · position-specific · a world model without a unified algorithm
Bayesian Induction
Bayesian Induction
prior → posterior update · the probabilistic reply to induction
Benoit Mandelbrot
Benoit Mandelbrot
Founder of fractal geometry · coined "fractal" 1975 · 1967 coastline paper · clouds are not spheres
The Bitter Lesson
The Bitter Lesson
general computation wins · don't hard-code what you think you found
Causal Contextualism
Causal Contextualism
do causal judgments depend on the evaluation context? three positions
Causal DAG
Causal DAG
chain / fork / collider · d-separation · the bridge to deconfounding
Causal Inner Product
Causal Inner Product
causally separable concepts are orthogonal under this inner product
Causal Intervention
Causal Intervention
do-operator · cut the original inputs · mini-surgery on the graph
Causal Models
Causal Models
Structural equations · intervention semantics · modularity
Causal Relata
Causal Relata
What kind of entity is a cause — and what kind is an effect?
Causation (Hume)
Causation (Hume)
Necessary connection is unobservable · habit is the real basis of causal inference
Chaos Engineering for Agents
Chaos Engineering for Agents
Fault injection · measuring reliability decay · simulating production stress
Chroma
Chroma
Open-source vector database · systematic Context Rot study · 18 models tested · RAG vs long-context empirics
Claude Agent SDK
Claude Agent SDK
Give-the-agent-a-computer · implicit-loop architecture · context compaction · long-running needs external tracking
Claude Code
Claude Code
Anthropic's official coding-agent CLI · three-tier tool approval · deny-first rules · five-level scopes · six permission modes
Claude Code Permission System
Claude Code Permission System
three tool tiers · allow/ask/deny rules · defense in depth
CLI-Anything
CLI-Anything
HKUDS 2026 · 29K stars · any software → agent-native CLI · 7-stage pipeline · CLI-Hub
CLI vs GUI Automation
CLI vs GUI Automation
two foundational paths for agent-driven software
Coastline Paradox
Coastline Paradox
smaller ruler → longer length · an intuitive gateway to fractal dimension
Codex
Codex (OpenAI)
OpenAI coding agent · codex-1 (RL-finetuned o3) · 6+ hour runs · zero-human million-line codebase · harness-engineering proof
Compositionality
Compositionality
meaning of a complex expression = parts' meanings × mode of combination
Constant Conjunction
Constant Conjunction
the empirical basis of Humean causation · all we actually observe
Constrained Decoding
Constrained Decoding
force illegal-token probability to zero · the mechanism behind structured outputs
Context Compression
Context Compression
the quality / compression-rate / readability trade-off across three strategies
Context Engineering
Context Engineering
find the smallest high-signal token set · curation, not accumulation
Context-Free Grammar (LLM)
Context-Free Grammar for LLM
JSON Schema → CFG → dynamic token mask
Context Management
Context Management
anterograde amnesia · working memory has to be programmed explicitly
Context Rot
Context Rot
systematic performance decay as token count grows
Counterfactual Theory
Counterfactual Theory of Causation
had c not occurred, e would not have · possible-worlds semantics
Custom and Habit
Custom and Habit
the general name for the associative principle · the real basis of causal inference
Cycle.js
Cycle.js
André Staltz · sources→sinks pure function · canonical fractal architecture · side effects pushed to driver layer
David Hume
David Hume
1711-1776 · Scottish philosopher · constant conjunction · problem of induction · custom · Kant's trigger
David Lewis
David Lewis
1941-2001 · counterfactual theory of causation · possible-world semantics · Influence theory 2000 · type causation = generalized token causation
Difference-Making
Difference-Making
the common core of multiple causal theories
do-Calculus
do-Calculus
Pearl's three rules — reducing causal queries to estimable observational expressions
Empiricism
Empiricism
all knowledge ultimately derives from experience · copy principle · method of definition
Error Cascade
Error Cascade
small upstream errors get amplified downstream · more than probabilities multiplying
Evaluator-Optimizer
Evaluator-Optimizer
separate generator and evaluator · GAN-style quality feedback loop
Factory.ai
Factory.ai
AI coding agent · anchored iterative summarization · 36K production-message compression eval · exposed artifact-tracking weakness
Falsificationism
Falsificationism
bold conjectures · try to falsify · science needs no induction
Feature Tracking
Feature Tracking
externalized machine-readable task list · blocks one-shotting and premature victory
Fork-Explore-Commit
Fork-Explore-Commit
OS-level exploration primitive · git semantics pushed down to execution · pick the best of parallel paths
Four Causes
Four Causes
material · formal · efficient · final · the starting point of causal pluralism
Fractal Architecture
Fractal Architecture
self-similarity · four rules · any subtree is a valid tree
Fractal Dimension
Fractal Dimension
quantifies space-filling · non-integer dimension · the exponent in Richardson's formula
Frontdoor Criterion
Frontdoor Criterion
identify causal effects via mediators when confounders are unobserved
Generation-Verification Loop
Generation-Verification Loop
AI generates · humans verify · verification speed sets efficiency · humans are the bottleneck
Gerard de Melo
Gerard de Melo
computational linguistics · trajectory-bias revealer · RANLP 2025 · Draft-Conditioned fix · instruct vs. base model divergence
Google
original A2A developer · ADK · Gemini · neutralized via Linux Foundation donation · full agent stack
Grue Problem
Grue Problem
new riddle of induction · same data supports contradictory conclusions · predicate projectibility
Guardrails
Guardrails
layered defense · runtime safety monitoring · optimistic-execution pattern
Harness Engineering
Harness Engineering
the control system wrapping the LLM · not a limiter — an enabler of reliable capability
Henry Kautz
Henry Kautz
University of Rochester · six-category neurosymbolic taxonomy · Third AI Summer · no Third Winter · Allen Newell Award
Herbert A. Simon
Herbert A. Simon
Nobel 1978 · Turing Award 1975 · bounded rationality · architecture of complexity · near-decomposability · PSSH co-author
Heuristic Search
Heuristic Search
how the physical symbol system solves problems · search volume ≠ intelligence
Hierarchical Systems
Hierarchical Systems
nested structure driven by interaction strength · watchmaker parable · an inevitable outcome of evolution
Implicit Loop Architecture
Implicit Loop Architecture
gather context · take action · verify work · repeat · behavior shaped by constraints
Indra's Net
Indra's Net
holographic mutual reflection · each node contains the whole · a philosophical discovery of self-similarity
Problem of Induction
Problem of Induction
no rational basis for "past so, future so" · Hume's fork
Inferential Theories of Causation
Inferential Theories of Causation
from regularity to inferential paths · causation characterized by its reasoning structure
Interventionist Theory
Interventionist Theory of Causation
causation = systematic answers to active manipulation · Woodward 2003
INUS Conditions
INUS Conditions
an Insufficient but Non-redundant part of an Unnecessary but Sufficient condition · Mackie 1965
James Woodward
James Woodward
University of Pittsburgh · Making Things Happen 2003 · interventionist theory of causation · modularity · type-level focus
Jeremy Howard
Jeremy Howard
fast.ai co-founder · Answer.AI founder · llms.txt standard · AI-education democratizer · top-down practice pedagogy
Jerry A. Fodor
Jerry A. Fodor
1935-2017 · Rutgers/MIT · language of thought · modularity · Fodor & Pylyshyn 1988 connectionism challenge
J. L. Mackie
J. L. Mackie
1917-1981 · INUS conditions · regularity theory's most important technical contribution · causal field · moral error theory
Judea Pearl
Founder of the causal revolution · Turing Award 2011 · UCLA CS Professor
Karl Popper
Karl Popper
1902-1994 · falsificationism · accepts the induction problem as unsolvable · corroboration vs. confirmation · Logic of Scientific Discovery
Knowledge Extraction and Fidelity
Knowledge Extraction and Fidelity
fidelity ≠ accuracy · describes what the network actually does
Ladder of Causation
Ladder of Causation
association · intervention · counterfactual · three irreducible rungs
LangGraph
LangGraph
LangChain · StateGraph explicit orchestration · the contrast paradigm to the implicit loop · Durable Execution · enterprise adoption
Language of Thought
Language of Thought
Mentalese · syntactically structured mental representations · the structural guarantee of systematicity
Lewis Fry Richardson
Lewis Fry Richardson
British mathematician and meteorologist · empirical discoverer of the coastline paradox · namesake of the Richardson Effect · pioneer of numerical weather prediction
Linear Representation Hypothesis
Linear Representation Hypothesis
features encode linearly in activation space · methodological basis of mechanistic interpretability
Linux Foundation
Linux Foundation
neutral open-source governance body · A2A protocol host · Kubernetes-style model · PyTorch Foundation
LLM OS
LLM OS
LLMs are not a utility · they are a software ecosystem · we are in the 1960s
LLM-OS Analogy
LLM-OS Analogy
LLM = CPU · Agent = OS kernel · Context = RAM · structural isomorphism
LLM Security
LLM Security
jailbreaks · prompt injection · data poisoning · the cat-and-mouse game never ends
LLM Training Pipeline
LLM Training Pipeline
pre-training → fine-tuning → RLHF · the canonical three-stage mental model
llms.txt
llms.txt
a root-path Markdown file · tells the LLM what this site is
Long-Running Agents
Long-Running Agents
work across sessions · each new session starts amnesic · shift-engineer analogy
Manus
Manus
general-purpose autonomous agent · 5-dimensional Context Engineering · prefill/decode 100:1 · logit masking
Max Tegmark
Max Tegmark
MIT physicist · Mathematical Universe Hypothesis · Future of Life Institute · spatiotemporal world model co-author
Maximilian Schall
Maximilian Schall
NLP researcher · named trajectory bias · draft-conditioned solution · RANLP 2025
MCP
Model Context Protocol
Anthropic open protocol · standard Agent ↔ tool interface · complementary to A2A · Mobile-MCP extension
Means-Ends Analysis
Means-Ends Analysis
GPS difference-driven search · recursive subgoals · the ancestor of CoT/ReAct
Mechanistic Interpretability
Mechanistic Interpretability
features · attribution graphs · CLT · hallucination mechanism · unfaithful CoT
Mediation Analysis
Mediation Analysis
TE · CDE · NDE · NIE · decomposing direct and indirect effects
MemGPT
Memory-GPT
UC Berkeley · virtual context management · LLM as OS · hierarchical storage · arXiv:2310.08560
Meta-Cognition in AI
Meta-Cognition AI
Monitor · Evaluate · Adjust · NSAI 5% blind spot · AlphaGeometry sole full coverage
Meta-Harness
Meta Harness
Session · Harness · Sandbox · brain-hand decoupling · OS virtualization analogy
Meta-Induction
Meta Induction
wMI strategy · object/meta levels · regret learning · bandit analogy
Near-Decomposability
Near Decomposability
Short-run independence · long-run aggregate coupling · frequency separation · Simon 1962 watchmaker
Necessary Connection
Necessary Connection
Hume's critique · three candidate sources ruled out · the mind's felt determination · Kant's reply
Neel Nanda
Neel Nanda
Google DeepMind · TransformerLens author · MATS mentor · OthelloGPT linearization · mechanistic interpretability
Nelson Goodman
Nelson Goodman
American philosopher · Grue problem · New Riddle of Induction · constructivist worldview · symbol theory
Neurosymbolic AI
Neurosymbolic AI
System 1+2 integration · existential/universal complementarity · Fodor-Pylyshyn challenge
Neurosymbolic AI Taxonomy
Neurosymbolic AI Taxonomy
Kautz's six types · coupling spectrum · from Symbolic_Neuro to Neuro[Symbolic]
No Free Lunch Theorem
No Free Lunch Theorem
generalization error 1/2 under uniform prior · Hume's induction problem, formalized · inductive bias is unavoidable
OpenAI
OpenAI
GPT series · Codex agent platform · Harness Engineering methodology · Agent Legibility · Structured Outputs
OpenHands
OpenHands
open-source multi-agent coding platform · CodeActAgent · SWE-EVO benchmark framework · performance = model × framework
Orchestrator-Workers
Orchestrator-Workers
dynamic task decomposition · distinct from predefined parallel · A2A cross-system extension
Othello World Model Hypothesis
Othello World Model Hypothesis
sequence training → board representation emerges · 99% accuracy · representation ≠ planning
Parallelization
Parallelization
Sectioning · Voting · predefined subtasks · confidence boost
Paul Smolensky
Paul Smolensky
Johns Hopkins · Harmony Theory · subconceptual level · tensor product representation · Optimality Theory phonology
Permission Modes
Permission Modes
six-mode autonomy spectrum · auto classifier · bypassPermissions nuclear option
Allow/Ask/Deny Rules Hierarchy
Permission Rules Hierarchy
Deny-First semantics · fixed evaluation order · Deny unoverridable across scopes
Pets vs Cattle
Pets vs Cattle
disposability · independent component recovery · brain-hand decoupling makes cattle-ification possible
Physical Symbol System Hypothesis
Physical Symbol System Hypothesis
Newell & Simon 1976 · necessary and sufficient for general intelligence · foundation of GOFAI
Possible World Semantics
Possible World Semantics
closest world · comparative similarity · non-backtracking interpretation · Lewis 1973
Preemption
Preemption
early · late · trumping · the core test-bed for counterfactual theories
Prefix Caching
Prefix Caching
KV cache reuse · 10× cost gap · stable static prefix first · at odds with compaction
Probabilistic Causation
Probabilistic Causation
probability raising · net vs path-specific effects · the preemption challenge
Probability of Causation
Probability of Causation
PN · PS · PNS · but-for legal standard · monotonicity identification
Probing Classifiers
Probing Classifiers
Linear/non-linear probes · High accuracy ≠ model uses it · Causal validation required
Problem Space
Problem Space
State-operator-goal · Representation sets difficulty · Newell & Simon 1976
Process Theories
Process Theories of Causation
conserved-quantity transfer · causal vs pseudo-processes · strong on preemption · blind to prevention
Prompt Chaining
Prompt Chaining
sequential steps · gate checks · trade latency for accuracy · simplest workflow pattern
Propositional Fixation
Propositional Fixation
McCarthy on neural-net limits · finite first-order logic · universal-quantification ceiling
Regularity Theory
Regularity Theory of Causation
Humean tradition · constant conjunction · INUS conditions · contemporary refinements
Reliability Decay
Reliability Decay
super-linear decay · positively-correlated errors · domain > model size · RDC slope quantifies decomposition payoff
Reliability Surface
Reliability Surface
3D evaluation · k consistency · ε robustness · λ fault tolerance · pass@1 is not enough
REPL for Agents
REPL for Agents
resident process state · undo as first-class primitive · dual mode · vs subcommand vs MCP
Rich Sutton
Rich Sutton
RL pioneer · The Bitter Lesson 2019 · general computational methods always win · keep stripping assumptions
Richardson Effect
Richardson Effect
shorter ruler · longer boundary · divergence as ε→0 · empirical origin of fractal dimension
Routing
Routing
classify, then dispatch to a specialized pipeline · separation of concerns · model selection · customer-support triage
Rule Circularity
Rule Circularity
justifying R with R itself · premise vs rule circularity · counter-induction's self-support problem
Scaling Laws
Scaling Laws
N×D predictable · no saturation in sight · theoretical basis of the Gold Rush · multi-step error cascades can cancel the gain
Settings Scope Hierarchy
Settings Scope Hierarchy
five-level priority · managed settings cannot be overridden · deny pierces every layer · auto-mode special rules
Software 3.0
Software 3.0
prompt is the program · natural-language coding · three generations · Karpathy 2025
Software Evolution Benchmark
Software Evolution Benchmark
SWE-EVO · single-step 72.8% → multi-step ~20% · release-notes input · error cascade exposed
Spatiotemporal World Model
Spatiotemporal World Model
Gurnee & Tegmark 2024 · R²=0.91/0.84 · linear probes · emergent space+time representations
SSM Hybrid Architecture
SSM Hybrid Architecture
Mamba-3 · Jamba 1/8 attention · linear complexity · long-context agent-inference optimization
Statistical Self-Similarity
Statistical Self-Similarity
natural fractals · statistical invariants across scales · basis of the Richardson effect · strict vs statistical
Structural Causal Model
Structural Causal Model
SCM ⟨U,V,E⟩ · unifies the three-rung ladder · structural equations ≠ regression · Pearl 2010 meta-framework
Structured Outputs
Schema→CFG→token mask · JSON mode vs Structured 11.97%→0.1% · hidden cost of trajectory bias
Subconceptual Level
Smolensky 1988 · microfeatures · three-level framework · neural-symbolic intermediate · distributed representations
Sumeru Mustard Seed
須彌芥子
no obstruction between great and small · Vimalakīrti Sūtra · Huayan school · precursor to self-similarity · Blake / Leibniz convergence
SWE-Bench
SWE-Bench
Princeton NLP · de facto standard for coding agents · GitHub issue repair · SWE-Bench Verified · 72.8% saturated
Symbol Grounding Problem
Harnad 1990 · dictionary loop · Chinese Room · three solution paths · LLM partial-grounding debate
System 1 vs System 2
Kahneman · Karpathy 2023 "LLMs are only System 1" · CoT/o1 as engineered System 2 · caution against dichotomy
Systematicity
Fodor & Pylyshyn 1988 · non-accidental linkage of abilities · central challenge to connectionism · LLM empirical test
Task Lifecycle
A2A state machine · submitted→working→4 terminal states · input-required human-in-loop · three tracking modes
Teleological Explanation
Aristotle · regularities demand explanation · non-psychological teleology · three causes unified · priority of the final cause
Thomas Bayes
Thomas Bayes
English Presbyterian minister & mathematician · Bayes' theorem · inverse probability · posthumous 1764 · probabilistic reply to induction
Token vs Type Causation
2×2 matrix · three positions on the relation · asymmetric existence cases · net vs component effects
Tool Design
format choice · minimal viable toolset · Poka-yoke · dual-mode dual-output · token efficiency
Trajectory Bias
Schall & de Melo RANLP 2025 · constrained-decoding cumulative perturbation · format-correct ≠ answer-correct · CRANE mitigation
Uniformity Principle
the future will resemble the past · Hume's dilemma · training distribution = test distribution · philosophical root of distribution shift
Virtual Context Management
MemGPT · OS virtual-memory analogy · RAM/disk hierarchy · page fault → memory retrieval
Wes Gurnee
Wes Gurnee
MIT · LLM interpretability · spatiotemporal representation R²=0.911/0.835 · sparse probing · ICLR 2024 co-author
Wesley Salmon
Wesley Salmon
philosopher of science · process theory of causation · mark transmission · causal-mechanical explanation · 1925–2001
World Models
LeCun AMI · predict the next state, not the next token · physical reasoning · LLM emergence evidence
02 Sources — 82
Each source is a structured digest of a paper, blog post, or design doc.
- Neurosymbolic AI: The 3rd Wave Garcez & Lamb (2020/2023)：神经符号 AI 第三波，Kautz 六类分类法详述，命题固着边界，忠实性 XAI 原则，Pearl 因果层级的神经符号实现
- Li et al. (2022/2023)：涌现的世界表征——OthelloGPT Li et al. (ICLR 2023)：OthelloGPT 奠基论文，非线性探针+激活干预验证棋盘表征的因果性，序列模型涌现世界模型的范式案例
- Emergent Linear Representations in World Models of Self-Supervised Sequence Models Nanda、Lee、Wattenberg (BlackboxNLP 2023)：OthelloGPT 线性化修正，Mine/Yours/Empty 坐标系使线性探针准确率从 ~75% 跃升至 ~99%，线性向量加法干预法，多电路假说
- Language Models Represent Space and Time Gurnee & Tegmark (ICLR 2024)：Llama-2 内部自发形成线性时空表征，空间神经元/时间神经元，世界模型基本成分的实证证据
- The Linear Representation Hypothesis and the Geometry of Large Language Models Park、Choe、Veitch (ICML 2024)：LRH 严格反事实形式化，因果内积统一子空间/测量/干预三种表征，LLaMA-2 实验验证
- From Words to Worlds: Compositionality for Cognitive Architectures Dhar & Søgaard 2024：LLM 组合性实证检验，规模提升组合性、指令微调降低组合性，ICML 2024 Workshop
- Neuro-Symbolic AI in 2024: A Systematic Review 2025 年 PRISMA 系统综述，167 篇论文，五维功能分类，元认知仅 5% 但最关键，AlphaGeometry 为唯一全谱案例，Kahneman System 1/2 自我批判
- Yuan & Søgaard (2025)：重访 Othello 世界模型假说 Yuan & Søgaard (ICLR 2025)：7 个 LLM 的 Othello 实验，99% 无监督接地精度，跨架构表征对齐 93–96%，世界模型假说的最强证据，但 2-hop 规划退化
- AgenticOS Workshop @ ASPLOS 2026 AgenticOS Workshop（ASPLOS 2026）：学术系统社区首次聚焦 agent OS 设计，12 篇论文涵盖资源控制、OS 接口、Skill OS、Fork-Explore-Commit、安全隔离
- Jamba: AI21's SSM-Transformer Hybrid Model Jamba：首个生产级 SSM-Transformer 混合架构，256K context
- AIOS: LLM Agent Operating System AIOS：Rutgers 团队的 LLM Agent OS，OS 六大模块映射到 agent 管理，context 快照恢复，2.1x 吞吐量提升
- On the Biology of a Large Language Model Claude 3.5 Haiku 内部机制研究：规划、幻觉、越狱、CoT 忠实性
- Building Agents with the Claude Agent SDK Claude Agent SDK 设计哲学：给 agent 一台计算机，隐式循环架构，gather-act-verify-repeat
- Building Effective AI Agents Anthropic 官方 agent 构建指南，workflows vs agents 分类，五种 workflow 模式，ACI 设计原则
- Circuit Tracing: Revealing Computational Graphs in Language Models 电路追踪方法论：跨层转码器、归因图、替代模型
- Claude Code 权限系统（官方文档） Claude Code 权限系统官方文档：三级工具分级、deny-first 规则语义、六种权限模式、五级作用域层次、权限与沙箱双层纵深防御
- Effective Context Engineering for AI Agents Anthropic context engineering 总论：注意力预算、context rot、just-in-time context、compaction/note-taking/sub-agent 三种长时策略
- Effective Harnesses for Long-Running Agents 长时运行 agent 的 harness 设计：initializer-coder 双 agent 架构、feature tracking、增量推进
- Emergent Introspective Awareness in Large Language Models LLM 内省能力研究：概念注入、思维检测、自我觉察
- Equipping Agents for the Real World with Agent Skills Agent Skills 开放标准：指令+脚本+资源的可组合打包，三层渐进式披露
- Harness Design for Long-Running Application Development GAN 式 generator-evaluator 架构、sprint contracts、context reset vs compaction、harness 随模型进化简化
- Harnessing Claude's Intelligence: 3 Key Patterns for Building Apps 三大模式：用已知工具、持续剥离假设、谨慎设边界
- Introducing Claude Opus 4.6 Claude Opus 4.6 发布：1M context、agent teams、adaptive thinking、effort 控制
- Scaling Managed Agents: Decoupling the Brain from the Hands Anthropic Managed Agents 架构：agent 组件虚拟化、brain-hands 解耦、session 外部化、meta-harness 概念
- Anthropic 结构化输出官方文档 Anthropic 官方结构化输出文档：output_config.format API、grammar-constrained sampling、文法缓存失效规则、可选参数复杂度模型、Extended Thinking 整合
- Tracing the Thoughts of a Large Language Model Anthropic 可解释性综述：AI 显微镜、归因图、十种行为的内部机制
- Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents 可靠性科学框架：RDC/VAF/GDS/MOP 四指标评估长时 agent
- The History of Pets vs Cattle and How to Use the Analogy Properly
- The History of Pets vs Cattle Pets vs Cattle 云计算类比：服务器可处置性的规范定义，基础设施设计的核心区分
- Mamba-3: An Inference-First State Space Model Mamba-3：推理优先 SSM，复数值系统、MIMO、经典控制论启发
- Context Rot: How Increasing Input Tokens Impacts LLM Performance Chroma 2025 研究：18 个前沿模型的 context rot 系统性测试，三种退化机制
- Claude Code 源码泄露：社区分析与内部机制研究（2026） 2026 年 Claude Code 源码泄露社区分析：cache 边界标记、autoCompact 机制、append-only JSONL 历史、子 agent fork-join KV cache、anti-distillation
- CLI-Anything：让所有软件都成为 Agent-Native CLI-Anything：HKUDS 开源项目，跨 8+ agent 平台的插件生态，生成 20+ 软件 CLI harness，CLI-Hub 自主发现注册中心，agent-native software 范式的代表实现
- Don't Break the Cache：长 Horizon Agent 任务的 Prompt Caching 评估 长 Horizon agent 任务 prompt caching 实证评估：500+ session，system-prompt-only 策略最优，45-80% 成本降低
- Evaluating Context Compression for AI Agents Factory.ai 压缩质量评估：三种策略对比，probe-based 功能质量测试，artifact tracking 难题
- Connectionism and Cognitive Architecture: A Critical Analysis Fodor & Pylyshyn 1988：系统性论证，联结主义为何无法替代经典符号架构，思维语言假说，实现解释建议
- Fractal Architecture（Anton Telesh, 2016） 分形架构：自相似组件、统一接口、递归组合，从 React/Elm/Cycle.js 提取的通用模式
- Fractal Buddhism: The Avatamsaka Sutra 华严经的分形解读：因陀罗网的 Cook 1973 经典描述、"无碍互入"即分形结构、学术资源
- Harness Engineering 为什么是 Agent 时代的"控制论"？ 控制论视角的 harness engineering：三次同构（调速器→K8s→harness），隐性知识编码
- Google A2A Protocol — 官方文档 Google/Linux Foundation 的 A2A 开放协议：Agent Card 发现、Task 生命周期、HTTPS/JSON-RPC、SSE 流式传输、OAuth 安全
- World Models Race 2026 世界模型竞赛：AMI Labs、Genie 3、Marble、Cosmos
- [1hr Talk] Intro to Large Language Models Karpathy 2023 标志性演讲：LLM OS 类比首次提出，三阶段训练流水线，scaling laws，System 1/2，LLM 安全攻防
- LLM = CPU, Agent = Operating System Kernel Karpathy 推文：LLM=CPU、Agent=OS Kernel，token vs byte，统计性 vs 确定性
- Software Is Changing (Again) — Andrej Karpathy YC AI Startup School 演讲 Karpathy 2025 YC 演讲：Software 1.0/2.0/3.0 分类法、LLM OS 类比、部分自主产品、autonomy slider、generation-verification loop
- The Third AI Summer: AAAI Robert S. Engelmore Memorial Lecture Kautz (AI Magazine 2022)：三个 AI 夏天历史框架，耦合谱权衡形式化，"不会有第三个冬天"论断，分类法权威原始来源
- LangGraph Documentation LangGraph：显式图编排框架，隐式循环的对比范式
- The /llms.txt file llms.txt 提案规范：Jeremy Howard 2024 年提出的网站 LLM 友好文档约定，格式规范、与 robots.txt/sitemap.xml 的关系、推理时 vs 训练时使用
- Manus：面向生产级 AI Agent 的 Context Engineering Manus 团队四次架构重建的 context engineering 实战：五维框架、KV cache 10× 成本优化、logit masking 保 cache、可恢复压缩原则、todo.md 注意力工程
- MemGPT: Towards LLMs as Operating Systems MemGPT：虚拟上下文管理，OS 层次化内存映射到 LLM context，长文档分析与多会话聊天
- I-JEPA: Image Joint Embedding Predictive Architecture I-JEPA：表征预测架构，世界模型基础，LeCun 愿景
- MIT Technology Review: Mechanistic Interpretability (2026 Breakthrough) MIT Tech Review 2026 十大突破：mechanistic interpretability
- Computer Science as Empirical Inquiry: Symbols and Search Newell & Simon 1975 图灵奖讲座：物理符号系统假说（PSSH）、启发式搜索假说、问题空间框架、弱方法与强方法、符号 AI 的定性结构律奠基
- Energy-Based Diffusion Language Models for Text Generation EDLM：能量扩散语言模型，非自回归文本生成
- Harness Engineering: Leveraging Codex in an Agent-First World OpenAI 零人工编码实验：agent legibility、repo 即知识系统、架构执行、熵管理
- Introducing Codex Codex 产品发布：云端 agent、AGENTS.md 引导、codex-1 模型
- Introducing Structured Outputs in the API OpenAI 2024 Structured Outputs 发布：JSON Schema → CFG → 约束解码，格式错误率从 ~11.97% 降至 <0.1%，符号规则约束神经生成的工程化实现
- A Practical Guide to Building Agents OpenAI agent 构建指南：设计基础、编排模式、guardrails、人类介入
- Unlocking the Codex Harness: How We Built the App Server Codex App Server 架构：threads/turns/items 会话原语、双向 JSON-RPC、多表面集成
- Unrolling the Codex Agent Loop Codex agent loop 内部机制：prompt 构建、推理流程、prompt caching、compaction
- jylin04 (2024)：OthelloGPT 学到了一袋启发式规则 jylin04 (MATS 6.0, 2024)：OthelloGPT 机制分析，空格检测电路逆向工程，表征准确≠算法统一，Bag of Heuristics 概念来源
- An Introduction to Causal Inference (Pearl 2010) Pearl 2010 因果推理权威综述：SCM 统一框架、do 算子、后门/前门准则、中介分析、因果概率、四步方法论
- The Hidden Cost of Structure: How Constrained Decoding Affects Language Model Performance Schall & de Melo（RANLP 2025）：约束解码的隐性代价，轨迹偏差命名与实证，11 模型实验揭示句法合规≠语义正确，指令微调模型 vs 基础模型的分歧行为
- The Elements of Operating-System Style
- The Elements of Operating-System Style Unix 哲学：统一思想塑造开发风格，"一切皆文件"和管道作为持久抽象
- ReliabilityBench: Evaluating LLM Agent Reliability Under Production-Like Stress Conditions ReliabilityBench：三维可靠性曲面 R(k,ε,λ)，生产级压力下的 agent 评估
- Schneier's Blog (2026-04 Snapshot) Schneier 博客快照：认知安全分类、加密法律、hackback
- Aristotle on Causality（亚里士多德论因果性） 亚里士多德因果性理论：四因说、aitia 概念、目的因辩护与解释优先性、假设必然性
- The Metaphysics of Causation (SEP) 因果关系形而上学（SEP）：五大因果理论（规则性/反事实/干预主义/概率/过程）、因果关系项、先占/预防/开关案例、因果模型
- 规则性与推断性因果理论（SEP） 规则性与推断性因果理论（SEP）：HRT→Mill→INUS→当代规则性、DN 模型→排名函数→Ramsey 测试→Kairetic、差异制造
- Counterfactual Theories of Causation (SEP) 反事实因果理论（SEP）：Lewis 1973/2000 理论、可能世界语义学、先占问题演化、结构方程框架、语境主义辩论
- 休谟哲学综述（斯坦福哲学百科） 休谟哲学综述（SEP）：因果性理论、恒常连结、必然联系、归纳问题、经验主义、道德情感主义
- SEP: The Problem of Induction 归纳问题（SEP）：休谟论证、齐一性原则、贝叶斯方案、证伪主义、Goodman grue 问题、No Free Lunch 定理、元归纳
- Herbert Simon —— The Architecture of Complexity（1962） Simon 1962 论文笔记：层级系统、近可分解性、钟表匠寓言、复杂性的描述与演化
- On the Proper Treatment of Connectionism Smolensky 1988：亚概念层理论，调和理论，功能性组合性，符号规则作为联结主义宏观描述
- 须弥藏芥子，芥子纳须弥 — 佛教典故的含义 须弥芥子佛教典故：智常禅师对话、阿含经米粒故事、大小无碍的哲学原理
- The Bitter Lesson
- The Bitter Lesson Rich Sutton 的苦涩教训：通用计算方法总是胜出，人类领域知识内建短期有效但长期阻碍进展
- SWE-EVO: Benchmarking Coding Agents in Long-Horizon Software Evolution Scenarios SWE-EVO：多步长 horizon 软件演进基准，48 任务平均 21 文件 874 测试，GPT-5.4 仅 25%，误差级联效应的核心证据
- Twitter/X Reactions: Managed Agents Launch Managed Agents 发布的 X/Twitter 社区反应：三种 harness 模式分类、Agent Sandbox Protocol 空白、平台吃掉中间件叙事
- Causal Model（因果模型）— Wikipedia 因果模型 Wikipedia 条目：结构因果模型、DAG、do 演算、因果之梯、Pearl 框架
- Coastline Paradox（海岸线悖论） 海岸线悖论 Wikipedia 条目：Richardson 效应、分形维数、统计自相似性、Mandelbrot 开创分形几何
- Physical Symbol System（Wikipedia） Wikipedia PSSH 条目：支持与挑战 PSSH 的论证全景（Dreyfus、Brooks、深度学习、符号接地问题、神经符号混合路径）