Claude Code Dynamic Workflows let Claude write its own JavaScript orchestration harness with subagents, parallelism, and per-agent model selection — live since May 28, 2026.
How the production agent stack is fracturing into distinct product layers — memory, skills, sandbox, eval, and harness — and what this means for 2026.
Four sources — Tsinghua papers, Melbourne ICL study, AgentFloor benchmark, deepagents-cli — converge: harness design drives a 6x model performance spread.
AHE framework lifts Pass@1 from 69.7% to 77.0%; harness-only changes yield 13–20% Terminal-Bench gains. Model upgrades no longer the only lever.
Stanford, Google/MIT, AHE, LangChain, and Unblocked all published this week: harness quality outperforms agent count as the primary agentic performance lever.
Curated AI insights — sent when there's something worth your inbox.