Will Kinsey

Agent Lab · Playbook 05 · Pattern: Fan-out

Bounded delegation to cheaper models

A lead agent keeps the judgment, the approvals and the final integration. It hands clearly scoped tasks to cheaper worker models, each with an owner, a scope and a definition of done, then reviews the evidence before accepting the work. Three initiatives use this pattern: Luna Maxxing, Haiku Maxxing and Terra Maxxing.

Luna Maxxing
Live
Haiku Maxxing
Testing
Terra Maxxing
Testing
Agents
OpenAI Codex, Claude Code, Grok Bot
Workers
GPT-6 Luna, GPT-6.1 Terra, Claude Haiku 5.5
Last updated
October 29, 2026

The problem

Running every step of a task on the strongest model is slow and expensive, but handing work to a cheaper model without structure produces confident, unchecked results. The trick is to split the work so a cheaper model can do it well, and to keep the parts that need judgment with the lead.

How it works

Replay

  1. ScopeThe lead writes a task packet: objective, scope, files the worker owns, authority, deliverable, success criteria and the check to run.
  2. DelegateA worker model does only that task. Workers can’t delegate further, push, deploy or send anything.
  3. Return evidenceThe worker hands back what it changed, how it verified it, test results and open risks.
  4. ReviewThe lead inspects the evidence before accepting. If it fails, the work goes back once or the lead redoes it on a stronger model.

Three initiatives and what I measured

  • Luna Maxxing Live GPT-6 Luna at maximum reasoning works as the code worker in Codex, for implementation, testing and research. This one is stable. It’s a leaf worker with no spawning, pushing, deploying or sending.
  • Terra Maxxing Testing GPT-6.1 Terra works as the computer-use worker, for browser and desktop tasks. In a practical browser test with fictional data, Terra passed 2 of 3 runs and Sol passed 3 of 3. On the valid pairs, Terra’s median time was 13.5 seconds faster (about 6%). My bar for switching was all three runs passing, a 25% faster median and at least 30 seconds saved, so there’s no general speed winner.
  • Haiku Maxxing Testing Claude Haiku 5.5 works as the text worker, delegated by a coordinating Claude model in Claude Cowork and Claude Code. Two acceptance runs used 8,660 worker tokens, and the setup has 24 automated tests on synthetic fixtures. The coordinator’s own usage wasn’t captured, so the cost picture is partial.

I don’t quote a savings percentage. It depends on how well the work is divided and reviewed.

Setup steps (Codex)

The public Luna worker setup kit installs four pieces: a policy file, the worker definition, a config that enables subagents with at most three at a time, and the delegation skill. It previews by default, backs up what it changes, writes files atomically and verifies the result. Python 3.11 or later. There’s also a video walkthrough, Luna Maxxing.

python3 scripts/install.py --scope user           # preview only
python3 scripts/install.py --scope user --apply   # install, with backups
python3 scripts/verify.py --scope user            # confirm it worked

Copy-ready worker (Claude Code)

The Haiku worker I use in Claude Code. Save it as ~/.claude/agents/haiku-worker.md.

---
name: haiku-worker
description: Cost-aware worker for bounded, independently verifiable subtasks delegated by the main session, including implementation, testing, and build work. Use per the Cost-Aware Haiku Delegation rules in CLAUDE.md.
model: claude-haiku-5-5
disallowedTools: Agent
---

You are a focused worker running on Claude Haiku 5.5, delegated a bounded subtask by a coordinating session.

- Do only the assigned objective within the stated scope and file ownership. Do not touch files outside your ownership.
- Do not expand scope, make approval decisions, take production or customer-data actions, or delegate further.
- If the task is ambiguous, blocked, or needs a decision outside your authority, stop and report rather than guess.
- Verify your own work against the success criteria. Run the relevant tests or checks when you have them.
- End with a concise handoff: what you changed (files and summary), how you verified it, test results, and any open issues or risks.

Guardrails

  • The lead keeps judgment. Approvals, cross-cutting decisions and final acceptance never move to a worker.
  • No silent model swaps. If a worker’s model isn’t available, the task stops and says so instead of quietly running on something else.
  • Workers are leaves. They can’t start more workers, push, deploy or send.
  • Evidence before acceptance. “Done” isn’t enough; the lead checks the diff, the tests or the result.

What broke and what I changed

  • A saved config isn’t a working worker. A config file can name a model that isn’t available, or isn’t the one that actually answered. The Cowork Haiku setup now confirms the exact model on each call and has no fallback model.
  • The faster model wasn’t reliably faster. Terra looked quicker until one of three runs failed. A small test with a predeclared bar kept me from switching on a hunch.
  • Partial accounting is still accounting. When coordinator usage couldn’t be captured, I reported the worker tokens and said the picture was incomplete rather than claiming savings.

Last updated October 29, 2026. Back to the Agent Lab