minssam.
Published on

Claude Code 2.1.292: Fine-Tune Subagent Cost and Speed with the Effort Parameter

The most painful part of multi-agent workflows is the invoice.

Breaking a big task into parallel agents speeds things up, but token costs rise in lockstep. Attaching a max-performance model to a simple file-listing task is like driving a sports car through a parking lot. Claude Code 2.1.292 solves this with a single effort parameter.

1. The Effort Parameter: A Five-Step Performance Dial

You can now pass effort when calling the Agent tool. Five levels are available: low, medium, high, xhigh, max.

Agent({
  description: "Summarize file list",
  prompt: "List files in src/ directory",
  effort: "low"
})

Agent({
  description: "Deep security audit",
  prompt: "Audit all auth logic against OWASP Top 10",
  effort: "max"
})

low is fast and cheap. max is thorough but slow and expensive. Pick the level that matches the stakes β€” you cut costs without sacrificing quality where it counts.

2. Priority Order β€” Which Setting Wins

When effort is set in multiple places, the hierarchy is:

  1. CLAUDE_CODE_EFFORT_LEVEL environment variable (absolute override)
  2. effort in the Agent tool call
  3. effort in the subagent definition frontmatter
  4. Parent session default (falls back to medium)

Team leads can set the env variable to cap the whole org. A developer coding max in their call still respects the env-variable ceiling.

3. Real-World Strategy: Cut Costs with a Two-Stage Pipeline

// Stage 1: Quick scan β€” filter irrelevant files
const scan = await Agent({
  description: "Fast file scan",
  prompt: "Classify changed files β€” separate test files from source files",
  effort: "low"
})

// Stage 2: Deep review β€” only the relevant set
const review = await Agent({
  description: "Code review",
  prompt: `Review these files in depth: ${scan.result}`,
  effort: "high"
})

Use low for tasks where mistakes are obvious and cheap to catch. Use max for security audits and final code reviews where a missed issue is costly.

4. plugin install --marketplace: One-Line Setup

The same release adds a --marketplace flag to claude plugin install.

Before (two steps):

claude marketplace add
claude plugin install my-plugin

Now (one step):

claude plugin install my-plugin --marketplace

If the marketplace is not yet registered in the current environment, this flag adds it automatically and then proceeds with the install. Especially useful in onboarding scripts and CI/CD pipelines.

5. Known Bug: --max-budget-usd Overshoot

Community reports indicate a rare bug in this version where the --max-budget-usd spending cap can be exceeded. In budget-sensitive environments, add extra monitoring or wait for the next patch.

Tips

  • Team cost control: Set the subagent default effort to medium in .claude/settings.json and require explicit high or max only where needed β€” this prevents surprise token bills.
  • EdTech use case: In an AI tutoring pipeline, a first-pass student submission review can run at low, while the final grading agent runs at high. Significant cost savings at scale.
  • CI automation: Adding --marketplace to your build scripts eliminates manual setup steps and reduces human error in fresh environments.
  • Incremental adoption: You don't have to change every agent. Unspecified agents default to medium, so you can optimize the most expensive agents first.

Sources:

Claude Code 2.1.292: Fine-Tune Subagent Cost and Speed with the Effort Parameter | MINSSAM.COM