Agents launch at full price

Status: searched 2026-07-31. Both blindnesses are documented on Claude Code itself; the unified thesis and the cheap-variant pattern are the deltas.

Claim: the model does not economize delegation. Left alone, it launches as many agents as the task shape suggests — without counting — and every subagent inherits the strongest available model, because nothing in the model’s sense of a good plan includes the bill. Fan-out size and lane models are governed from outside or not at all.

The two blindnesses

Count-blindness. The model sizes a fan-out by the task’s structure, not its budget: one lane per file, one per dimension, one per question, however many that makes. My repo carries a hard cap (at most twelve agents per workflow) because sizing was never reasoned — and the cap rule records that the owner has questioned too FEW lanes as well as too many. It is not that the model over-launches; it is that the number is never a decision.
Tier-blindness. Subagents default to the expensive model. The model never volunteers “this lane is grep-shaped, use the small model” — the economizing thought is simply absent, the same cost-blindness that prices a 500k-token demonstration at zero (the demonstration reflex, which is this finding’s demo-shaped special case).

What works

Everything that works is imposed, none of it volunteered:

A tier ladder written as caps, enforced by lint. My repo’s delegation rule names a model tier per task kind (frontier only for load-bearing verification and synthesis; mid-tier for implementation; small models for grep-shaped fan-out), makes every tier a CAP rather than a floor, requires named owner consent for the top tier — and a repo check FAILS any saved workflow lane whose model is unset. The fail-closed lint is the load-bearing part: the default it overrides is “inherit the expensive one.”
Cheap variants of expensive skills. I keep a second version of my deep-research skill with the same harness shape — fan out searches, fetch, adversarially verify, synthesize — where only scoping and synthesis run on the session model; search runs on the smallest tier and fetch/verify on the middle one. Same output shape, a fraction of the cost. The point is who authored it: the human. The model, asked to research something, reaches for the full-price version every time; the economical variant exists only because I built it and told sessions when to use it.
A price on the ask. Fan-outs above a size threshold are proposed with lane count and models named before dispatch, and demonstrations are purchases (the demonstration reflex).

Prior art

Verdict: PARTIAL. Both blindnesses are documented, including on Claude Code itself. Anthropic’s own multi-agent research writeup (2025) reports “early agents… spawning 50 subagents for simple queries” and hand-written scaling rules as the fix. A practitioner post, “The Subagent Tax” (Systima, 2026-07-22), measures both on Claude Code directly — fan-outs reaching 41, then 102, then 415 agents, and Explore agents “inherit the parent model rather than defaulting to Haiku.” On the academic side, BAGEN (arXiv:2606.00198) finds frontier models “consistently over-optimistic” about their own spend, and BAMAS (arXiv:2511.21572, AAAI 2026) sets agent count and per-agent tier jointly with an external optimizer, stating that “existing work rarely addresses how to structure multi-agent systems under explicit budget constraints.” pilotfish’s per-role model pins remain closest to the tier-ladder half. The delta: the two blindnesses as one structural property (nothing in the model’s plan prices tokens), the thesis that every working delegation economy is human-imposed and machine-enforced, and the human-authored cheap-variant-of-an-expensive-skill as the practical fix — none found stated as such.