Claude Duet 用 Opus 5.5 编排多个常驻小助理

damkin @damkina7原推发布于 2026.10.22

在 X 上打开原推

WHAT THE ACTUAL F*CK?! Anthropic just took direct aim at OpenAI Dots and Grok Bot with Claude Duet The single chat window is officially dead OpenAI gave you a pulsing dot. Anthropic dropped an army of autonomous workers Under the hood, Claude Duet is powered by Claude Opus 5.5, orchestrating a decentralized fleet of persistent micro-workers: 1. The Death of the Monolithic Context Window Instead of dumping research, file management, and writing into one bloated chat session that suffers from prompt drift, Duet isolates tasks into dedicated worker identities. Your "AI Research" bot keeps 41 of 960 sources, while "Computer Use Bot" autonomously organizes local directories in the background. 2. Native Computer Use Directly on Mobile Substrates This is not another API wrapper. The inclusion of background Computer Use means agents actually execute actions (sorting Downloads, archiving 83 newsletters, staging pull requests) rather than just generating suggestions 3. The "Waiting On You" Deterministic Gate The fatal flaw of autonomous agents has always been runaway hallucinations. Duet introduces a hard verification gate: "Never posts without asking" and "2 waiting on you" The agent runs the full compute loop independently, but execution halts at the commit barrier until you hit approve 4. The Gamified Trust Layer (The Trojan Horse) Disguising production-grade autonomous agents as customizable 8-bit pixel bots (custom hats, shades, and shells) is a deliberate UX masterstroke. Users resist complex CLI orchestrators, but readily manage a personal fleet of Tamagotchi-style specialists The Anatomy of an Agent Failure: The Monolithic Chat Cascade: Single Prompt -> Context Bloat -> Hallucinated File Edits -> Unchecked Execution -> Catastrophic Error The Claude Duet Substrate: Isolated Specialist -> Opus 5.5 Background Execution -> Human Gate ("Waiting on You") -> Verified State The Operational Takeaway: • Stop trying to build a single mega-agent that does everything. Production systems require task-specific isolation • Never give an autonomous worker write-access without an explicit approval boundary OpenAI gave you a glowing circle. Anthropic just gave you an army of autonomous workers Deploy your autonomous fleet on Claude Duet: ↳ https://claude.com/product/claude-duet

这条是引用着另一条写的：x.com/claudeai/status/2112984664256241837