# Subagent Configuration Inherits From the Parent With No Override ## Claim In current agentic coding harnesses (Codex confirmed, Claude Code's own behavior unverified as of this writing), a spawned subagent inherits the parent session's model and reasoning-effort level with no exposed control to set it independently — so raising your own effort level multiplies cost by however many subagents get spawned, not just once. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: Theo (t3.gg), "gpt-5.6-sol without hitting limits" (X/Twitter thread, 2026-07-14). > [!quote] > "gpt-5.6-sol will ALWAYS spawn subagents with same model and reasoning level as the parent instance." Theo ties this directly to why Codex's "Ultra" mode is "broken" — Ultra's high reasoning level propagates to every subagent it spawns, compounding cost with no way to cap it. His workarounds are all indirect: lower your own reasoning level before letting the model spawn subagents, add an anti-spawning rule to a global instructions file, or use an undocumented config flag to inspect spawn metadata. None of them let you set a *different, cheaper* config for children while keeping a capable parent. ## Scope & conditions - Confirmed for Codex's gpt-5.6-sol as of 2026-07-14; not independently verified for Claude Code's own Agent/Workflow dispatch — worth checking directly rather than assuming the same limitation applies. - This is a harness/tooling constraint, not a fundamental property of subagent orchestration — a harness that exposed per-subagent model/effort selection would eliminate the problem entirely. - Applies specifically to *reasoning-effort tiers within the same provider/model family* (Codex's max/ultra), not to explicit cross-model routing (Claude Code choosing Haiku vs Opus for a subagent), which several vault notes already document as a deliberate, controllable pattern. ## Links - [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — because contrasts: that note's cost-saving pattern (expensive model plans, cheap model executes) assumes you *can* set the cheap model on the worker; this note names a harness where you currently can't, when "cheap" means a lower reasoning tier of the same model rather than a different, explicitly-selected model - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — because applies-in: a hidden cost multiplier specific to subagent-spawning harnesses ## MOC - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]]