Seven Mental · 心智七篇
中 EN
← Knowledge Atlas · Entity

Anthropic

AI 安全公司，Claude 开发者，本 wiki 最主要参考来源
实
ENTITY · ANTHROPIC · AI safety company · Claude · interpretability · agent infrastructure

Anthropic

AI safety company, maker of the Claude series — one of this project’s primary references

Anthropic is one of the primary references for this wiki: the workflows-vs-agents taxonomy for agentic systems, the augmented-LLM building block, the five workflow patterns, ACI design principles, and the harness-engineering methodology all draw on its official docs. Its mechanistic interpretability work is a frontier in the field — MIT Technology Review listed it as one of the 10 Breakthrough Technologies of 2026.

Four core domains
Agent frameworkClaude Agent SDK (formerly Claude Code SDK): “give the agent a computer”, implicit-loop architecture, built-in context compaction
InterpretabilityCircuit Tracing attribution graphs · Biology of an LLM (systematic study on Claude 3.5 Haiku) · emergent introspective awareness — MIT TR 2026 top-10 breakthrough
Managed AgentsBrain-hands decoupling: reasoning and execution separated, the first production realization of the meta-harness concept, borrowing from OS virtualization
Structured outputsGrammar-constrained sampling — constraints apply only to final output; Extended Thinking reasoning segments stay unconstrained
Flagship models
Claude Opus 4.6
2026 flagship — first Opus-class 1M-token context, agent teams + adaptive thinking
MCP stewardship
Anthropic-led Model Context Protocol — the standard interface for agent tool access
→ Claude Code · Claude Agent SDK · Mechanistic InterpretabilityAnthropic Official Docs (2024-2026)

Anthropic

AI 安全公司，Claude 系列模型的开发者。

与本 wiki 的关联

Anthropic 是本项目最主要的参考来源之一。其关于 agent 构建的官方文档定义了许多核心概念：

  • Agentic systems 的 workflows vs agents 分类
  • Augmented LLM 作为基础构建块
  • 五种 workflow 模式的系统化梳理
  • ACI 的概念和设计原则
  • Harness engineering 的实践方法论

相关实体

  • Claude Agent SDK — Anthropic 的 agent 开发框架
  • MCP — Anthropic 主导的模型上下文协议

可解释性研究

Anthropic 在 mechanistic interpretability 方面是业界领先者：

Claude Opus 4.6

Opus 4.6 是截至 2026 年的旗舰模型，首个 Opus 级 1M token context，agent teams、adaptive thinking、effort 控制等产品特性。

Claude Code

Claude Code 是 Anthropic 的官方 AI 编码 agent CLI，将 Claude 嵌入开发工作流。其权限系统是生产级 agent 权限管理的参考实现——三级工具分级审批、deny-first 规则语义、五级作用域层次、六种权限模式，以及与 OS 沙箱的双层纵深防御。

Managed Agents

Managed Agents 是 Anthropic 的托管式 agent 服务，将 agent 组件（session、harness、sandbox）虚拟化为稳定接口。核心架构是 brain-hands 解耦——推理与执行分离，每个组件可独立故障恢复和替换。这是 meta-harness 概念的首个生产实现，直接借鉴了 OS 虚拟化的设计方法。

Context Engineering

Effective Context Engineering 是 Anthropic 对 context engineering 的系统性论述——从 prompt engineering 到 context engineering 的演进、注意力预算与 context rot、just-in-time context 策略、长时任务的三种策略（compaction、structured note-taking、sub-agent 架构）。

Structured Outputs

Anthropic 提供结构化输出 API（output_config.format），使用 grammar-constrained sampling（文法约束采样）保证模型输出符合 JSON Schema。底层机制与 OpenAI Structured Outputs 同类，但 Anthropic 对文法作用域、缓存失效规则、可选参数复杂度模型有更细化的文档化说明。关键设计：grammar 约束只作用于直接输出，不影响 Extended Thinking 的推理区段，让模型保留思考阶段的自由度。

References

  • sources/anthropic_official/building-effective-agents.md
  • sources/anthropic_official/effective-harnesses-long-running-agents.md
  • sources/anthropic_official/harness-design-long-running-apps.md
  • sources/anthropic_official/building-agents-claude-agent-sdk.md
  • sources/anthropic_official/tracing-thoughts-language-model.md
  • sources/anthropic_official/introducing-claude-opus-4-6.md
  • sources/anthropic_official/circuit-tracing-methods.md
  • sources/anthropic_official/biology-large-language-model.md
  • sources/anthropic_official/emergent-introspective-awareness.md
  • sources/anthropic_official/harnessing-claudes-intelligence.md
  • sources/anthropic_official/effective-context-engineering-for-ai-agents.md
  • sources/anthropic_official/claude-code-permissions.md
  • sources/anthropic-managed-agents.md