Cost patterns -- B12, B13, B14, B15, B16, A12, R5

Token economics adds seven catalogue entries. Five at TIER 2 design patterns (Bxx), one at TIER 3 architectural pattern (Axx), one refactor pattern (Rxx).

TIER 3 architectural

A12 GRADIENT WORKFLOW

Heavy planner-class front stage; mid-class implementer fan-out across N similar items; lightweight reviewer/triage back stage. The agentic analog of Tiered Architecture: pay heavy class only where reasoning is genuinely required.

flowchart LR
  P[planner: scope + decompose] --> I1[implementer 1]
  P --> I2[implementer 2]
  P --> IN[implementer N]
  I1 --> R[reviewer: coherence + synth]
  I2 --> R
  IN --> R

Pays off when N >= ~4 mid workers per planner call (heavy front amortized across fan-out width). For N < 4, flat single-class design is structurally simpler with comparable cost.

Anti-patterns: INVERTED GRADIENT (cheap front -> expensive bulk), ROUTER-AS-PLANNER (front stage doing real planning on trivial class).

TIER 2 design

B12 MODEL ROUTER

A trivial-class classifier dispatches each request to one of N downstream branches, each bound to a different role class. Pays off when traffic is heterogeneous and the cheapest meeting-capability branch dominates traffic.

Rule: router cost must be < 5% of the cheapest downstream call, else routing is the dominant cost.

Anti-pattern: ROUTER-AS-PLANNER. If the router does the work instead of dispatching, it cannot be trivial class.

B13 CACHE-AWARE PREFIX

Place all stable content (persona, rubric, tool catalogue, system prompt) BEFORE the explicit cache breakpoint. Variable content (user task, payload, retrieved spans) goes AFTER. Cache discipline is BOOLEAN per turn — one invalidator and the entire prefix re-bills at input rate.

Invalidators to audit: timestamped persona (”## Current date:”), mid-session model switch, mid-session MCP tool catalogue churn, mid-session effort change (extended thinking on/off).

Largest no-tradeoff lever in the catalogue. A design that correctly applies B13 typically halves input billing on repeat turns.

B14 PROMPT THRIFT

Remove rationale, hedging, and worked examples from always-loaded prefix. Move to references/ files loaded on demand. Coding discipline applied per-module.

Trigger: prefix budget over 80% AND recent edits added prose (rationale/hedging) rather than capability.

B15 TOOL SUBSET

Expose only the tools this module needs. Catalogue bleed (every MCP server’s full tool list loaded into every session) inflates the stable prefix proportionally to the wider toolchain, not the module’s actual needs.

Frugal stance mandates. Apply proactively to skills with tight prefix budget, even if MCP catalogue is small today.

B16 EFFORT GOVERNOR

Declare the reasoning-effort level (e.g. Claude extended thinking on/off, OpenAI reasoning effort low/med/high/xhigh) once per session. Mid-session effort changes invalidate cache (B13 invalidator).

Stance binding: frugal -> minimum/none; balanced -> low-med per role; quality -> high-xhigh on planner/critic; unbounded -> unrestricted.

Refactor

R5 COST PRUNE

Applied to existing designs that exceed budget. Step 1: OBSERVE the trace (do not guess) — capture a real run’s input-prefix bytes per turn, total turns, total output tokens, observed cache hit ratio. Then apply the cost-shape matrix to identify the dominant bucket and the right refactor (B13, B14, A12, etc.).

Anti-pattern: COST-OPTIMIZED-BY-VIBES — applying B12+B13+B14+B15 reflexively without measuring first.

Selection rule

The cost-shape matrix in pattern-tradeoffs.md Section 10 is first-match-wins. Do not stack all seven patterns reflexively. Each pattern names a structural choice with a deterministic cost shape; pick the one whose shape matches the workflow.

Anti-patterns introduced by this layer

The cost-economics addition created its own catalogue of failure modes worth naming:

• CLASS-UNIFORM GRAPH — every module on the most-capable model. R5 trigger.
• INVALIDATOR LEAK — timestamp / model switch / tool churn inside the cached region.
• COST-OPTIMIZED-BY-VIBES — pattern stacking without observing a baseline trace.
• HARDCODED MODEL NAMES — binding to “Claude Opus 4.7” instead of the planner role class; designs age out within 6 months.
• ROUTER-AS-PLANNER — the B12 router doing real planning on trivial-class capacity.
• INVERTED GRADIENT — A12 with the cheap class on the front stage and the expensive class on the bulk fan-out.
• BUDGET-DRIVEN PROMOTION — promoting to heavier class because the budget allows it, not because capability requires it.