In plain terms: Instead of working with one AI assistant and waiting while it finishes each task, I run several at the same time — one writing code, one testing, one writing documentation — the way a manager runs a small team. My job shifts to assigning the work, checking it, and making the final calls.

I run Claude Code and OpenAI Codex as parallel sub-agents: implementation, debugging, and documentation happening concurrently instead of serially. This note is the operating model, condensed.

The loop

  1. Context & plan — give each agent clear goals, relevant context, constraints, API assumptions, and an explicit completion contract
  2. Orchestrate & build — break work into bounded tasks, connect the right tools, coordinate implementation across the stack
  3. Verify & debug — inspect outputs, compare payloads, run checks, trace failures, iterate until acceptance criteria pass
  4. Human review — stay at every consequential decision; document tests, tradeoffs, and handover notes

What parallelism is actually for

Not typing speed — waiting. While one session runs a build or a test suite, another is drafting the migration, a third is writing the PR description. The human becomes the scheduler and reviewer, which is the job that was always worth the salary.

Boundaries that keep it sane

  • One session per bounded task, not one session per project — sessions that sprawl accumulate stale context and drift
  • Durable knowledge goes in the repo (CLAUDE.md, specs, docs) — never rely on a session “remembering”
  • Every merge is human-gated. The commit trail carries the AI co-author trailers, so the split of labor is auditable, not vibes

What I learned

The productivity isn’t in the agents typing faster — it’s in never sitting idle while something runs. My job stopped being “write the code” and became “keep three streams of work unblocked and review what comes back,” which is a management skill, not a typing skill. The surprise: the boundaries (one session per task, knowledge in the repo, human-gated merges) matter more than the number of agents.

Related: Write the plan before the code, Letting an AI hunt for the bugs humans skim past