GitHub Copilot CLI now delegates to subagents 23% less often when it shouldn't. A production A/B test of the new orchestration policy, shipping in version 1.0.42, recorded a 27% reduction in search tool failures, an 18% reduction in edit tool failures, and a 5% improvement in wait time at P95, with zero quality regression. The fix is live for 100% of Copilot CLI traffic. Run /update in your terminal to get it.
The root cause was eager delegation. The main agent was handing off narrow, fully-described tasks to subagents that then re-searched the repository from scratch, accumulated stale file paths, and stalled while the main agent idled. The new orchestration policy draws a hard line: focused single-file edits stay in the main agent, subagents are reserved for broad exploration, cross-cutting investigation, or work that is genuinely parallelizable. The change sounds simple. The path to it was not.
What makes this worth reading in full is the methodology. The team used LLMs to analyze full agent trajectories at scale, identify the orchestration bottleneck, translate the diagnosis into a revised policy, validate offline, then confirm with a production A/B test before shipping. That feedback loop, from telemetry signal to policy change to shipped improvement, is described in detail. If you build agentic systems and care about how to measure and correct orchestration overhead without regressing quality, the process documented here is the actual artifact.
[READ ORIGINAL →]