Reduce latency and review-loop overhead in subagent workflows

[tjgreen42]

Describe the feature or problem you'd like to solve

Subagent-driven implementation and review workflows can feel extremely slow. Agent startup, task handoffs, and repeated review/fix/re-review cycles often take minutes per round trip, making modest repository changes take much longer than expected.

Proposed solution

Reduce subagent startup and handoff latency, reuse context more efficiently between fix and review rounds, and collapse redundant review passes when prior findings have been addressed. Show clearer elapsed-time and current-work indicators so users can distinguish useful background work from a stalled workflow.

Example prompts or workflows

1. Implement a small configuration option plus regression coverage.
2. Run an implementer followed by a code-review agent.
3. Apply one review fix and re-review only the changed portion.
4. Complete a documentation-only task without another expensive full-context handoff.

Additional context

Observed with Copilot CLI 1.0.83 on Linux. The main concern is end-to-end latency and the number of sequential agent round trips, not model answer quality.

[tjgreen42]

Clarification from the reported experience: the concern is specifically Copilot CLI orchestration and tool-call implementation overhead, not that the review workload was unusually extensive. A modest final review reached roughly 100 tool calls and 11+ minutes despite prior task-level reviews and a passing full test suite. Tool and agent round trips need substantially lower latency regardless of call count.

[tjgreen42]

User-impact clarification: this is not merely an annoyance or benchmark concern. The orchestration latency is preventing productive completion of normal engineering work. The reporter can complete comparable workflows effectively with Claude Code but cannot get work done efficiently with the current Copilot CLI stack. End-to-end task completion time should be treated as a core usability and competitiveness issue.

[tjgreen42]

Magnitude clarification: the reporter estimates Copilot CLI is roughly 10× slower than Claude Code for comparable engineering workflows. This is a user-reported end-to-end productivity difference, not a small latency regression. Please evaluate complete task time—including agent startup, tool calls, handoffs, and review loops—rather than isolated model response latency.

[tjgreen42]

Personal impact, posted explicitly at the reporter’s request: the approximately 10× productivity difference and inability to get work done with this stack are severe enough that this experience makes the reporter consider quitting Microsoft. This should be treated as a serious developer-experience and retention concern, not routine performance feedback.

[tjgreen42]

Competitive-impact clarification: when developers cannot reliably finish normal work because of Copilot CLI orchestration overhead—and report alternatives as roughly 10× faster—this becomes a strategic competitiveness risk, not just a UX defect. The reporter believes the current implementation risks losing developer adoption and falling behind competing agentic coding tools.