Second Brain

subagent-driven-dev-scale-pattern

2 min read

subagent-driven-dev-scale-pattern

superpowers:subagent-driven-development at 46+ task scale — batch mechanical, focus full-cycle on logic

R-TaskFlow shipped end-to-end in a single session (2026-04-17, 55 commits, 90 tests, live at r-taskflow.arjtech.in) via superpowers:subagent-driven-development executing a 46-task plan. Naive literal interpretation of the skill (implementer + 2 reviewers per task) would have been ~140 subagent dispatches × 2-3 min each = 8+ real hours plus massive controller context burn.

Pattern that actually works at scale

  1. Batch tightly-coupled mechanical tasks into single implementer dispatches. Scaffolding tasks, multi-file similar-shape work (compose files, icon generation, 5 shadcn primitive installs, 4 Server Actions with near-identical structure). Verify inline with Bash. Skip separate reviewer subagents.

  2. Reserve full 3-subagent cycle (implementer + spec-reviewer + code-quality-reviewer) for logic-heavy single tasks. Auth actions, task state transitions, cron jobs, authorization-rule-dense work, security-adjacent code.

  3. When a subagent reports DONE_WITH_CONCERNS, decide inline whether concerns are scope (ship) vs correctness (fix-pass). Don’t auto-dispatch another reviewer.

  4. Keep TodoWrite milestone-level (~10 items) not task-level (46 items) after the first couple tasks. The plan file is the source of truth for granularity; TodoWrite is the user-facing progress beacon.

  5. Deploy at milestone boundaries, not after every commit. R-TaskFlow deployed ~7 times across the session (after M0, M1+2, M3, M4+5, M6, M7, M8+9+10). Each redeploy via ./scripts/deploy.sh takes ~90s and verifies /api/health.

Result

  • 55 commits / 90 tests / 6 real hours
  • Typecheck + lint clean at every milestone
  • Zero context pollution — controller stayed under compaction threshold
  • Zero regressions

Guidance for the skill itself

The skill’s base text implies “implementer + 2 reviewers per task” literally. Add a “Scale Guidance” addendum that codifies batching + inline-verify for mechanical tasks. See evolution-backlog 17-Apr-2026.

Graph View