Complex tasks need understanding before action. The combo workflow separates exploration from implementation:
- Plan mode — Read-only exploration. Map dependencies, understand architecture, design the approach.
- Explore subagent — Isolate verbose discovery from large codebases. Return a summary to preserve main context.
- Direct execution — Implement with full understanding. Modify files based on the confirmed plan.
Each phase has a specific role. Plan mode cannot implement (read-only). Explore subagent cannot implement (returns summaries). Direct execution should not explore deeply (wastes context, risks premature changes).
The Session Log Pattern
A successful combo workflow session:
| Phase | Duration | Files read | Files modified | Rollbacks |
|---|---|---|---|---|
| Plan mode | 15 min | 12 | 0 | — |
| Direct execution | 8 min | 0 | 4 | 0 |
Read-only exploration, clear transition, focused implementation, zero rollbacks, all tests pass. The exploration built understanding; the execution applied it precisely.
The Premature Switch Anti-Pattern
A developer used Plan mode to explore an agent orchestration pattern, then switched to direct execution. During implementation, they discovered 2 additional services that needed changes, causing context overflow and session restart.
The plan phase did not explore deeply enough. Switching to direct execution before identifying all affected components led to the exact rework the combo workflow is designed to prevent.
The rule: do not switch to direct execution until the plan phase has fully identified the scope, dependencies, and approach. If you discover new dependencies during execution, return to Plan mode — the workflow is iterative.
Iterative, Not One-Shot
Session logs from a multi-day refactoring:
- Session 1 (Plan): read 18 files, identified 3 service boundaries
- Session 2 (Direct): modified 7 files across 2 services, all tests pass
- Session 3 (Direct): modified 2 files, 2 rollbacks on service C → switched back to Plan mode
Session 3 shows healthy iteration: service C had dependencies the initial plan did not reveal. Returning to Plan mode to re-explore is the correct response — not pushing through with more rollbacks.
Even thorough planning can miss dependencies that only surface during implementation. The ability to return to Plan mode is a strength, not a failure of initial exploration.
When Each Mode Leads
Plan mode leads when: architecture is unknown, multiple approaches exist, dependencies are unclear, or the change spans many files across services.
Direct execution leads when: the solution is known, the scope is identified, and the target files are clear. A known bug fix, a well-specified feature in a familiar codebase.
Explore subagent leads when: the codebase is large (500+ files) and thorough exploration would exhaust the main session’s context.
Mixed-Mode Sessions
Two CI issues in one pipeline: a known typo blocking deployments (direct execution) and a new canary deployment stage requiring infrastructure research (plan mode).
Fix the urgent typo immediately with direct execution. Then use plan mode to research the infrastructure for the canary stage. Each issue gets the mode matching its characteristics.
Combining with Explore Subagent
For a 2,000-file codebase: Plan mode starts the exploration, but reading hundreds of files would fill the context. Delegate broad discovery to the Explore subagent (isolated context), get the architectural summary back, then switch to direct execution for implementation. The main session has 88% of its context preserved for edits.
One-liner: Plan mode maps what needs to change, Explore subagent does it without filling context, direct execution implements the confirmed plan — and the workflow is iterative, not one-shot.