execplan-improve-subagents

To improve an ExecPlan using parallel AI subagents, run a two-wave workflow where eight dedicated reviewers audit the original plan, then the parent agent rewrites it in place. This delegates evaluation to specialized subagents while keeping the parent as sole author, reducing drift.

No, subagents cannot directly edit plan files during a code review workflow. They operate under a strict read-only policy, meaning the parent agent performs all in-place rewrites to produce the final, cohesive ExecPlan.

The best way to audit a software engineering plan with multiple reviewers is a two-wave subagent workflow. Wave one spawns eight parallel reviewers on the original plan, and wave two spawns two closure reviewers on the rewritten draft to ensure code-grounded improvement.

To preserve key sections like Decision Log when rewriting an ExecPlan, the parent agent performs in-place rewrites while explicitly maintaining core sections including Progress, Surprises & Discoveries, Decision Log, and Outcomes during the subagent-driven improvement process.

Yes, the execplan-improve-subagents workflow aligns with Ousterhout design principles. It structures the plan improvement process to reduce cognitive load and complexity by coordinating specialized subagents for parallel evidence gathering and evaluation.

You should use a two-wave subagent workflow for plan improvement when you need code-grounded, parallel-evidence-driven evaluation and explicit closure checks. It applies when auditing a plan with parallel reviewers or delegating improvement tasks to subagents.