Post by Murali Koujala

Anthropic Just Absorbed Your Orchestration Layer You changed one string. claude-opus-4-7 → claude-opus-4-8. Nothing broke. The benchmarks improved. You shipped. Something else happened while you were reading the release notes. Anthropic absorbed your orchestration layer. 🔄 What Dynamic Workflows Does Dynamic Workflows — now in Claude Code for Enterprise, Team, and Max — doesn't just spawn subagents. It: → Plans the work → Dispatches tens to hundreds of parallel subagents → Has agents adversarially refute each other's findings → Verifies outputs before anything reaches you That plan → dispatch → verify → report loop? You used to build that. The planner, the fan-out, the state management, the retry logic — that was your harness. Now it's a platform primitive. The proof: Bun ported from Zig to Rust. ~750k lines. 99.8% test-suite passing. Eleven days. 🧱 What's Still Yours → Context engineering — what goes into the 1M window → Your evals — your definition of correct → Your domain — what the platform cannot see The orchestration layer was always undifferentiated heavy lifting. The teams who treated it that way are in a better position today. 💡 Prompts "I maintain a bespoke orchestration harness for multi-agent work: [describe it]. Given that Dynamic Workflows now handles planning, fan-out, adversarial verification, and checkpointing as a platform primitive, identify which parts of my harness are now redundant versus which parts still encode genuine domain-specific value that the platform cannot replicate." #ClaudeCode #Anthropic #AIEngineering #GenerativeAI #AgenticAI #LLMOps #SoftwareEngineering