Microsoft whitepaper evidence: Swap out the Harness, and the same Claude performs noticeably worse.

Microsoft’s November 7 technical whitepaper on Agent execution systems reported that swapping out the Harness behind Copilot Studio while keeping Claude Opus 4.8 raised complex-task pass rates by 22.1 percentage points, with model calls per task nearly doubling but converted input-token consumption falling 21%. The result lands just ten days after Microsoft said it would extend its unified Copilot Harness across multiple products and sell hosted access via its cloud platform, and follows Kimi’s November 5 launch of Hosted Agents under a “Harness as a Service” model, DeepSeek’s August MIT-licensed Harness release, and earlier developer-facing Harness offerings from OpenAI and Anthropic. Harness—the runtime layer that lets models call tools, persist task state and resume long-running work—is becoming a standalone product, with Anthropic’s Managed Agents and OpenAI’s Codex and Agents API among the commercial offerings. Research cited in the article shows Harness choice can reshape model rankings: the same Kimi K3 scored 66.7% task pass rates on Codex, 63.3% on DeepSeek Harness Creator and 56.7% on Kimi Code, while a Meta-Harness experiment lifted text-classification accuracy 7.7% and cut context-token use to a quarter. Automated Harness optimization carries risk, with one case dropping SWE-bench performance from 94% to 32% after tuning on new tasks. Watch whether Harness becomes a durable paid layer for model companies and how pricing evolves, given that model inference remains the primary revenue line.