# Orchestration That Terminates by Code Beats Termination by Model Choice ## Claim When a multi-agent system's stopping condition is written in code as fixed phases rather than left to the model's judgment, total cost and runtime stay bounded no matter how much work a single phase generates. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for any agentic system where token cost scales with agent count and turn count, and where sub-agents inherit the parent's reasoning-effort setting with no independent cap. - Breaks down if the fixed phases themselves are too coarse-grained to bound the work inside a single phase — e.g. a phase that fans out to an unbounded number of sub-agents with no per-phase cap just relocates the unboundedness one level down. ## Evidence Codex's "Ultra" mode spawns sub-agents that each inherit the parent's maximum reasoning effort with no override, and the whole system "ends when the agents decide to end, not when the code is completed" — a documented case of one user exhausting a 5-hour usage limit in 20 minutes. The contrast case is Claude Code's workflow model, where phases are defined in code and terminate on a coded condition rather than an open-ended agent decision. ## Links - [[AI & the Future of Knowledge Work MOC]] — why the stop condition is the most dangerous judgment to leave to the model - [[3. Resources/Permanent Notes/Subagent Configuration Inherits From the Parent With No Override]] — because extends; that note documents the inheritance mechanism this claim identifies as the root cause of unbounded cost. - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — because applies-in; both notes describe cost structures in agent systems that compound silently unless a design constraint bounds them. - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — because applies-in; this is a general design principle for controlling agent-system spend. ## Tests / falsifiers - Compare total token cost and wall-clock runtime across matched tasks run once with a fixed-phase (code-terminated) workflow and once with an open-ended sub-agent-judgment system, holding the underlying model and task constant. - Find a counterexample: a model-judgment-terminated system that reliably self-bounds cost (e.g. via an internal budget-awareness instruction) without a coded stopping condition — this would weaken the claim's generality. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Podcasts/I can't believe they released this]] — Theo (t3.gg) walkthrough of Codex Ultra's sub-agent cost behavior.