What changed
Microsoft released a managed Copilot harness in Foundry Agent Service with an agent loop, sandbox, skills, authenticated tools, approvals, files, code execution and built-in subagents. A new Copilot Studio paper reports 10.8- and 22.1-point pass-rate gains over its Standard harness on two tested models while cost-weighted tokens fell 15% to 21%.
Why it matters
The paper supplies measured evidence that harness design can change task success, cost and recovery even when the model is held constant. That makes orchestration, caching and tool-loop behavior first-class evaluation variables rather than invisible product plumbing.
What to do
Recreate a representative long-running task on the Standard and Copilot harnesses with the same model, tools and grading. Record pass rate, cache hit rate, cost-weighted tokens, failed-tool recovery, wall time and duplicate external actions before choosing the managed preview.
Caveat
The comparison is Microsoft-authored and not independently replicated. Task sets, the unnamed competitor and full production sampling are not exposed in the announcement; the harness remains public preview, and applications still need idempotent tools and durable external checkpoints.