Summary

The author runs a four-subagent Claude Code session on a .NET refactor, then stops mid-session to audit the workflow's costs. The expected culprit — duplication overhead from parallel agents — turns out not to be the biggest problem. Status-polling that pulled full agent transcripts into the orchestrator's context introduced a large, compounding cost that dwarfed the parallelism tax. This leads to a core realization: tokens are a one-time expense, but context pollution taxes every subsequent decision in the session. The author concludes that subagents are primarily valuable not for speed but for keeping disposable intermediate reasoning out of the orchestrator's main thread. Standing rules codified in CLAUDE.md should state missing facts rather than encode bureaucratic approval processes.

Key Insight

In long-running multi-agent workflows, the critical scarce resource is not tokens or parallelism slots but the quality of the orchestrator's working memory — context pollution compounds across every subsequent turn, making what the orchestrator chooses to carry forward more consequential than how many subagents it runs.

Spicy Quotes (click to share)

• 7

  Tokens are spent once. Context shapes every decision that follows.

• 8

  The tax I went looking for was never on the subagents. It was on the orchestrator, in what it chose to carry forward.

• 6

  What matters is not how much room there is, but how much of what's sitting in that room is worth the model's attention.

• 4

  The orchestrator is the only part of the system that accumulates understanding across a long session.

• 3

  Tasks that need the same mental model should usually stay together. Splitting them just forces multiple agents to rebuild the same understanding from scratch.

• 4

  Exploration, repeated file reads, failed approaches, and noisy intermediate reasoning are meant to stay in worker contexts and never make the trip back to the main thread.

• 7

  I wasn't really improving governance. I was just adding another ritual.

• 4

  Before adding a line to a standing instruction file, ask whether a reasonably competent orchestrator would make the right decision once it knew the one missing fact.

Tone

reflective, analytical, pragmatic