## 1. Article Identity ### Jiten Oswal ### Medium (CodeToDeploy), 11 February 2026 ### *The Architecture of Scale: A Deep Dive into Anthropic's Sub-Agents* ## 2. Core Argument ### Anthropic's Research system implements the Orchestrator-Worker pattern at production scale ### Sub-agents run in isolated context windows; the orchestrator never inherits their full traces ### Multi-agent with this pattern outperforms single-agent by ~90% on complex tasks (Anthropic internal eval) ## 3. The Anthropic Setup ### Lead Researcher (often Claude 4.5 Opus) analyses the query and delegates ### Worker sub-agents run faster cheaper models (Sonnet, Haiku) on specialised tasks ### CitationAgent post-processes for source attribution ### Quote: "each sub-agent operates in an 'isolated context window,' preventing context degradation in the main thread" ## 4. Communication Protocol ### Sub-agents "report back only the summary or result" to the orchestrator ### No peer-to-peer sub-agent channel ### Maintains clean context window for the orchestrator, even when sub-agents do extensive tool-calling internally ## 5. Spawning Heuristics ### Parallel research across multiple API endpoints — natural fit ### Code review and verification — natural fit ### Complex multi-step tasks requiring isolation — natural fit ### Simple queries stay in the orchestrator (no spawn) ## 6. Cost and Routing ### Multi-agent systems consume ~15× more tokens than chat (consistent with Boroumand's number) ### Model routing matters: Haiku for linting/simple logic, Sonnet for coding, Opus for architectural reasoning ### Implication: production cost depends as much on routing discipline as on the pattern itself ## 7. Why It Matters ### Concrete implementation reference for the convergent pattern ### Quantitative anchor for the "+90% on complex tasks" claim (Anthropic internal eval) ## 8. Source - https://medium.com/codetodeploy/the-architecture-of-scale-a-deep-dive-into-anthropics-sub-agents-6c4faae1abda - Accessed: 2026-05-23