Workflow economics: the router moves from the request to the task

Sources (2026-09-26): Ken Huang, "The Next AI Moat Is Workflow Economics" (Agentic AI, RSS + Gmail); a practitioner critique of request-level routing (@kunchenguid, X home feed); OpenRouter's Fusion announcement (Gmail, OpenRouter newsletter); Business Analytics Review, "The SaaSpocalypse has a winner" (Gmail). Raw: raw/rss/2026-09-26-agentic-ai-the-next-ai-moat-is-workflow-economics.md, raw/gmail/2026-09-27-newsletters.md.

TL;DR

Four independent pieces on one day make the same argument from different seats. Cheap tokens did not make AI cheap: they raised appetite for agent steps (plan, act, verify, retry, repair), so the bill moved from the prompt to the workflow. Ken Huang argues the durable moat is "workflow economics": deciding which jobs deserve premium reasoning, with routing lanes, promotion rules, caps and a KPI of cost per successful outcome instead of cost per call. A practitioner critique sharpens where that routing has to live: a bare request is too little signal to judge difficulty ("how does this work" is trivial in a one-file repo and hard in the Linux kernel), and switching models mid-session throws away the prompt cache, so request-level routers can cost more than using the best model throughout. The fix is task-level routing inside the harness. OpenRouter, a request-level gateway, shipped the opposite tool the same week: Fusion fans one prompt out to up to eight models and has an analyst report agreements, contradictions and blind spots. Business Analytics Review names the profit pool: whoever controls the router between fixed-price contracts and falling compute cost.

flowchart LR
  T[Task] --> S{Strong model<br/>scopes the task}
  S -->|cheap subtask| C[Small or<br/>decision model]
  S -->|hard subtask| F[Frontier model]
  S -->|high stakes| P[Fusion panel<br/>up to 8 models]
  C --> K[Cost per<br/>successful outcome]
  F --> K
  P --> K
  R[Request-level<br/>gateway] -.loses context,<br/>breaks cache.-> F
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class T input
  class S decision
  class C,F,P aux
  class K output
  class R warn

Key claims

Cheap units, expensive totals. Every drop in model cost invites more steps; the unit price falls while workflow appetite rises faster (Huang, Figure 1).
Most margin leaks sit outside the model call: fan-out, oversized context, repeated retries, and cleanup after a run went on longer than the job was worth.
Governance is cost control. The same boundary that limits risk (permissions, stop rules, validators before planners) limits waste.
Request-level routing has two structural problems: missing context and prompt-cache loss on model switches.
Fusion is overkill for tactical prompts, by OpenRouter's own description; the calling model decides per request whether to invoke it.

Relation to prior wiki pages

Names the frame the concept page has been building. llm-routing recorded on 09-26 that error correlation between tiers, not calibration, decides whether a cascade gains anything (Jev vs LLM rubric judges (09-26)). Fusion's "blind spots" field is a product built on the same insight: diverse models are valuable only where their errors differ.
The cache argument has prior data. The 09-23 price-war page found cache reads fell about 60% in the Opus 5.5 and GPT-6 Sol/Luna release (price war, cache reads (09-23)). Cheaper cache reads raise the penalty for a router that breaks the cache.
Harness research agrees on where cost lives. HarnessTax (09-22) and harness choice costs, not success (09-17) found harness decisions move cost more than success rate.

Open question

Nobody has published a head-to-head of request-level versus task-level routing with prompt-cache costs counted. That is the missing experiment.