Parallel Agents: Three Ways to Run Concurrent Work

Three Parallelism Models

• **Claude Code** — Multiple Agent tool calls in one message run concurrently, each with its own context. Our `/dx-step-verify` runs lint + secrets + architecture checks in parallel, cutting verification time by ~60%.
• **Copilot CLI /fleet** — Spawns N parallel subagents, one per subtask. Results converge back. Think MapReduce for code tasks. Great for homogeneous work like code review across many modules.
• **VSCode Chat #runSubagent** — Creates a context-isolated sub-task. The subagent works independently, returns results to your main conversation. Your context stays clean. Best for one-off heavy research or delegation.
• **Don't parallelize dependent tasks** — If step 2 needs step 1's output, run them sequentially. Parallel agents can't see each other's work.

When to Use Which

1. 1 **Match the model to the task type**
   • › Heterogeneous (different tools) — Claude Code multiple Agent calls
   • › Homogeneous (same task, N targets) — Copilot CLI /fleet
   • › Delegation (one-off research) — VSCode Chat #runSubagent
2. 2 **Try /fleet** — In Copilot CLI, run `/fleet "summarize each file in src/components/"` to see parallel agent execution in action.
3. 3 **Look for parallel opportunities** — Check if any of your sequential skill steps could run in parallel. Independent verification checks (lint, secrets, architecture) are prime candidates.
4. 4 **Keep dependencies sequential** — Map your skill steps as a dependency graph. Only parallelize steps with no data dependencies between them.