Routing inside the coding agent: skill granularity and the context-price cliff

Sources: arXiv 2610.31461 Adaptive Skill Granularity (Zhang, Jiang, Wang, Zhou; SJTU, Warwick, Kuaishou), Kurate cs.AI weekly #20 and a DAIR.AI Academy pick (DAIR page, surfaced via @dair_ai); practitioner thread on Claude Code subagent sizing (@Sprytixl) · raw: raw/kurate/2026-10-25-cs-ai.md, raw/twitter/feed/2026-10-25-morning-ranked.json Topic: routing, agent harness, cost control

TL;DR

Two routing decisions inside one coding agent. Adaptive Skill Granularity treats "how big a step to take" as a runtime policy: primitive tool calls, a macro skill, or a macro that falls back to primitive steps when it fails. A feature-based controller picks the level from task structure, execution state and earlier macro failures; a finite-state executor makes the fallback an explicit transition rather than an ad hoc retry. In a 12-task routing study a learned router reaches 100% success where a hand-written heuristic fails, and macro-to-micro fallback recovers controlled macro failures that fixed granularity leaves unresolved. The context cliff thread claims a cheap model's price jumps 5x past a context threshold, so an orchestrator should fork workers before they cross it; the poster reports one job at $6.15 as a single agent versus $0.85 split across nine. The linked Claude Code model-config page does not show that tier, so treat the numbers as the poster's.

Two routers, one agent loop

One picks the step size; the other keeps each worker under a price threshold.

flowchart LR
  T["Task<br/><small>coding request</small>"] --> C["Granularity router<br/><small>tool, macro, or hybrid</small>"]
  C -->|fits| M["Macro skill<br/><small>cheap when it fits</small>"]
  C -->|unclear| H["Macro then micro<br/><small>explicit fallback</small>"]
  M -->|fails| H
  H --> W["Worker size check<br/><small>fork before cliff</small>"]
  W --> O["Verified result<br/><small>verifier-defined success</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class T input
  class C,W loop
  class M,H core
  class O exit
  linkStyle 3 stroke:#e03131,stroke-width:2px

Blue is input, amber are routing decisions, purple is execution, green is the result; the red arrow is the failure path.

Key points

Fixed macro execution is efficient when the skill fits and brittle under mismatch: the classic routing trade-off, now at the skill level.
Evaluation is deterministic with verifier-defined success, total cost, macro adoption, recovery, route match and transition correctness.
The study is small (12 tasks, controlled failures). Kurate's AI rating is low (3.5/10); the DAIR pick is the stronger quality signal.
The context-cliff pattern generalizes the 10-23 "Opus plans, Haiku reads" note on the routing page: add a size constraint per worker and route by projected context, not only by task difficulty.

Relation to prior wiki pages

Extends RouteWeaver (10-24, one policy picks the execution mode) down one level: the mode router picks single-round vs agentic; this router picks tool vs macro inside the agentic mode. Routing-by-subtask is now n≥7.
Connects to Anthropic's "fold subagents into the main loop" guidance (10-24, summary): skills are the unit that replaces subagents, and this paper is the first to treat skill execution as a recoverable policy.
Open: neither the router reward nor the cliff heuristic prices prefix-cache loss when a worker forks, the gap flagged on 10-23/24.