# A Smart Orchestrator Should Delegate to Cheaper Executors ## Claim You get the most from a frontier coding agent by letting it orchestrate while cheaper models handle token-heavy and computer-use subtasks. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds when a cheaper model has generous quota and adequate capability for the delegated subtask — routing overhead has to stay smaller than the savings. - Breaks when the delegated work is taste-sensitive or requires the orchestrator's own context to do well; routing overhead or quality loss then dominates. - This is a workflow-design claim, not a claim that any specific model pairing (Fable/GPT-5.5) is the optimal one — the pairing is one instance, not the generalization. ## Evidence Theo's ~5.5-hour, ~$150 Fable 5 run steered GPT-5.5 via Codex CLR for token-heavy and computer-use subtasks rather than doing everything inside the frontier model directly, keeping both subscriptions under 40% weekly usage while shipping roughly a month of backlog. The pattern only works because Fable's value is autonomy horizon (orchestration, decomposition to sub-agents) rather than raw per-step intelligence — the orchestrator's job is deciding what to delegate, not doing everything itself. ## Links - [[AI Agent Safety MOC]] — cost and safety both improve when the expensive model directs rather than executes - [[3. Resources/Permanent Notes/Agency Growth Is Healthy Only When Margin and Delegation Scale]] — because extends: delegation only pays off when the delegate has real capacity margin, whether that delegate is a person or a cheaper model. - [[3. Resources/Permanent Notes/Bounded Action Spaces Not Better Models Make Agents Safe to Run]] — because applies-in: safe delegation depends on scoping what the executor is allowed to touch, not on trusting its judgment. - [[3. Resources/Permanent Notes/Direct Client Access Needs Billing Guardrails To Protect Margin]] — because contrasts: that note is about protecting margin from uncontrolled human access; this one is about protecting margin from uncontrolled token spend — same discipline, different resource. - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — because applies-in: model-routing-by-cost is a direct tokenomics decision. ## Tests / falsifiers - Compare total cost and defect rate for a matched task run entirely on the frontier model versus run with the frontier model orchestrating a cheaper executor for token-heavy subtasks. - Look for cases where routing overhead (context handoff, re-verification of the cheaper model's output) erases the cost savings — that would bound the claim's scope rather than falsify it outright. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/I Finally Get Fable 5]] — reports the orchestration pattern and the ~$150/5.5-hour run it's based on.