AI 原生开放教材

外观

Claude Code Ultimate Guide工程手册·英文原文 ↗原件 ↓仓库 ↗

10. Sources ​

Official Anthropic Sources ​

  1. Introducing Claude Opus 4.6 Anthropic, Feb 2026 Official announcement of Opus 4.6 and agent teams research preview

  2. Building a C compiler with agent teams Anthropic Engineering, Feb 2026 Technical deep-dive: git-based coordination, autonomous C compiler case study

  3. 2026 Agentic Coding Trends Report Anthropic, Jan 2026 Production metrics: Fountain (50% faster), CRED (2x speed)

Community Sources ​

  1. Claude Opus 4.6 for Developers: Agent Teams, 1M Context dev.to, Feb 2026 Setup instructions, workflow impact table, read/write trade-offs

  2. The best way to do agentic development in 2026 dev.to, Jan 2026 Integration patterns: Claude Code + plugins (Conductor, Superpowers, Context7)

Community Tools ​

  1. Claude Agent Teams UI Open-source desktop app (Electron + React + TypeScript) for managing Claude Code agent teams. Kanban board with real-time task tracking, code review diffs, cross-team communication, deep session analysis, and context monitoring. 100% free, runs locally.

Practitioner Testimonials ​

  1. Paul Rayner LinkedIn Post Paul Rayner (CEO Virtual Genius, EventStorming Handbook author), Feb 2026 Production usage: 3 concurrent workflows (job search app, business ops, infrastructure)

  2. IFTTD ep 311 "IA Agentique" Samy Lastmann (CTO, Smart Tribune), 2024 Production multi-agent system: micro-agents replacing one mega-prompt. Each agent has a single task and the corresponding tools. Scalability, explainability, and measurability all improve when agents are split by responsibility. Not every agent in the pipeline needs an LLM; deterministic steps (regex, BM25, SQL lookups) belong to code, not to a reasoning model.

  3. IFTTD ep 361 "Pourquoi le RAG n'est pas mort" Guillaume Laforge (Developer Advocate, Google Cloud), 2025 Introduced "harness engineering" as the discipline of managing context flow across a multi-agent pipeline. Orchestrating isolated sub-agents is as much a context architecture problem as it is a task decomposition problem: each agent receives only the slice of context it needs, and the harness controls what gets passed where.

  4. IFTTD ep 341 "Bilan 2025" Quentin Adam (CEO, Clever Cloud), 2025 Multi-model orchestration pattern: Claude decomposes a task and orchestrates Gemini (larger context window) to execute large codebase rewrites that would exceed Claude's window. Planning and task formulation stay with the stronger reasoner; bulk context ingestion goes to the model built for it.

  5. IFTTD ep 346 "IA & DevX" Jocelyn N'takpe (Head of Engineering & Architecture, ManoMano), 2025 At scale (250 engineers, 400+ microservices): a dedicated platform team maintains skills and rules encoding internal architecture conventions (database connection patterns, Kafka message encoding, naming standards). Agents are treated as new developers who need onboarding. Agent-facing documentation is now better maintained than Confluence. MCP stack in production: Serena (semantic code search via LSP), Playwright, Context7.

  6. Devoxx, 2026: "The Spec Paradox" Max Sumrall, 2026 Restricting an agent's task too tightly degrades output quality nearly as much as giving it no specification at all; broader, more ambitious goals sometimes pull more creative results out of the agent than a narrow brief does. A documented case at Picnic showed the same pattern at team level: perceived productivity rose sharply in the first weeks of AI-assisted development, then declined once code shipped faster than it could be reviewed, forcing repeated rework on the same features.