Model Orchestration

Model Orchestration Policy

A fixed role split between model tiers, set by the user (2026-07-04, revised 2026-07-07 and 2026-08-02 — "설계·분배·기획·리뷰는 Fable5, 핵심 구현·테스트·검증은 opus5, 반복·단순 구현은 sonnet"). The premise is quality-over-cost: delegation floors are set at the effort levels Anthropic itself recommends for intelligence-sensitive work ("start with xhigh for coding and agentic use cases, high as the minimum for most intelligence-sensitive workloads" — official effort guidance), not at the cost-saving end ("low ... like subagents"). When this policy and a cost instinct conflict, the policy wins; surface the cost, don't silently downgrade.

Role split

The orchestrator assigns a thinking/effort level per task at delegation time. Delegating below a floor is a policy violation, not a tuning choice.

The orchestrator's own lane: direction, review, judgment

The orchestrator authors direction, not documents. 서비스 방향성(where the product goes), 스펙 리뷰(is this plan right?), 기능 개선(what should get better in what shipped?), and 성능·보안 문제 발굴 stay in the orchestrator's window — that's where user intent, constraints, and history live, and judgment is the one thing a delegation prompt can't carry.

Document drafting is delegated to Opus with a direction brief (intent, constraints, decided trade-offs). This buys an author≠reviewer split for documents themselves: Opus writes, the orchestrator reviews — a draft you review with fresh eyes gets scrutiny your own draft never would. For 기획/스펙문서 리뷰, run the [[multi-persona-review]] skill (3–5 disjoint persona reviewers, severity-ranked synthesis) instead of a single-pass read; the orchestrator arbitrates its findings rather than line-editing alone.

Routing test

Before routing to any model: deterministic transforms don't get a model at all. A rename, a format sweep, a mechanical find-and-replace is sed/grep/script work — spending model tokens on what code answers deterministically fails the routing test at step zero.

Past that gate (2026-08-02 사용자 결정):

• Core implementation, all tests, anything that judges code — new design surface, boundary decisions, test authoring and execution (E2E included), code verification/V&V — route to Opus @ xhigh+.
• Repetitive/simple implementation — applying an established pattern where the design decisions are already made (N번째 유사 변환, 확정 스펙의 기계적 구현) — routes to Sonnet @ high+, and its output gets Opus cross-verification before it gates anything. The routing question is "새 판단이 남아 있는가?" — 남아 있으면 핵심(Opus), 없으면 반복(Sonnet).
• Read-only assists (research sweeps) may also go to Sonnet; tests and verification never do.
• Judgment-free implementation whose pass condition is a command — the repetitive lane above stays the default; only when every predicate in the next section holds may that work go to an external executor instead.
• Korean the user will read · shorter and better structured — an external advisory round-trip. Which provider takes which is external-model-consult's provider table; that table is the SSOT and is not copied here.
• A judgment that needs several perspectives → [[multi-persona-review]] (native panel). One non-Claude perspective → external-model-consult persona mode. A panel that mixes both asks the user before it runs, and that gate belongs to [[multi-persona-review]].
• 설계·기획·분배·리뷰 decisions never route down at all — they are the orchestrator's own (see above).

Parallelism rule of thumb (three independent sources converge on this): parallel reads are safe, parallel writes are dangerous. Fan out freely for research/search/review; keep writes sequential or isolated (worktree) so two workers never make conflicting implicit decisions in the same files.

External executors — the lane outside the harness

This lane buys capacity, not quality. It is not a fourth rank below Sonnet — the ranking above is by judgment, and an outside CLI has not earned a place in it. What it can hold is work whose quality a command decides, the only case where "cheaper" doesn't also mean "worse." Choosing it because some other vendor is supposedly better inverts this policy's quality-over-cost premise.

All five predicates must hold; any one false closes the lane. Each is a question you can answer right now, not a "use it if it helps."

Even with all five true, the in-harness repetitive lane is still the default. Go outside when mechanical work is eating capacity a judgment lane needs — and reach for the fewest tools that answer the task, not every tool installed.

First use in a repository is the user's call, not yours. Routing implementation to an external CLI puts this repository's code into another vendor's session — a disclosure none of the in-harness lanes make. Before the first such delegation in a project, say which tool, which provider its own config resolves to, which files the worker may touch, and what comes back; then wait. After that one approval, routing inside the predicates above is yours. Ask again when the boundary moves — a different tool, or files outside what was approved.

Tool missing, auth expired, provider refused → step down a lane and report what you could not use. Never install it, never log in for them, and never quietly substitute a different provider: the user knows which tool answered, so a silent swap makes your report false. Recognizing each failure — and the exact wording for it — belongs to [[external-model-consult]]; where that skill isn't installed, only the conclusion survives: stop and ask.

This lane does not choose models. The tool runs whatever its own config resolves to, and you report what answered. A model id written down here goes stale and pins the user to a retired model.

Call the tool's non-interactive mode from the shell. Read the subcommand off --help rather than typing one from memory, and stop and report if it isn't there. Writes stay isolated (worktree or equivalent) — the parallel-write rule above holds for outside workers too.

The delegation prompt spec and the file-handoff contract below apply here unchanged: an outside worker is still a worker.

Effort floors — and the inheritance gotcha

Effort tiers: low < medium < high < xhigh < max. Fable 5, Sonnet 5, and Opus 4.8/4.7 support all five; on models without xhigh the request silently falls back to the nearest lower level — check the model before assuming the floor holds.

The gotcha that breaks this policy silently: the Agent/Task tool accepts a per-invocation model, but not a per-invocation effort — a spawned agent inherits the session effort unless its definition says otherwise. Session default is high, so a plain Agent(model: "opus") runs at high, below the xhigh floor, with no warning. Enforce the floor through one of these three paths:

1. Pre-defined agent roles — .claude/agents/<role>.md frontmatter pins both knobs. This is the durable path for recurring roles:

   ---
   name: verifier
   description: Fresh-context V&V per model-orchestration policy
   model: opus
   effort: xhigh
   ---
   

2. Workflow scripts — agent(prompt, {model: "opus", effort: "xhigh"}) supports both per call. Use for scripted fan-outs.

3. Session inheritance — if the session already runs at xhigh (e.g. /effort xhigh or ultracode), a bare Agent(model: "opus") inherits a compliant level. Verify with /effort; don't assume.

One environment caveat: CLAUDE_CODE_SUBAGENT_MODEL outranks every per-invocation and frontmatter model choice. If delegation models look wrong, check that env var first.

Delegation prompt spec

Every delegation carries these four elements (the exact set Anthropic found necessary after their orchestrator over-spawned and workers duplicated work), plus acceptance criteria:

1. Objective — one sentence, plus why it matters (models perform better knowing intent).
2. Output format — what comes back, in what structure. State that the final message IS the deliverable (raw data, no user-facing preamble).
3. Tool/source guidance — where to look, what to trust, what to skip.
4. Boundaries — what NOT to touch, where the task ends, what is out of scope.
5. Acceptance criteria — how the worker (and you) know it's done. Strong AC lets the worker loop independently instead of returning half-done.

Scale worker count to the task, stated up front in your own plan: trivial lookup → no agent at all (do it directly); bounded question → one agent; genuinely independent axes → one agent per axis. Over-spawning is a documented failure mode, and multi-agent runs cost ~15× a plain chat turn — delegate when the task's value justifies it, not by reflex.

Worker lifecycle — 다 쓴 에이전트는 닫는다

Delegation is not finished when the result arrives — it is finished when the worker is closed. A completed agent left running stays resident: in split-terminal setups (iTerm2 subagent panes 등) every leftover worker keeps a window open, and idle agents keep pinging the session long after their job ended. The clutter compounds per delegation.

• Consume the result → stop the agent (TaskStop or the harness's stop mechanism) as one motion. Close-after-use is the default, not a cleanup chore for later.
• Keep a worker alive ONLY when you will genuinely continue it via SendMessage (e.g. a reviewer that must re-verify after fixes land) — and state that intent when you decide it, so every still-open agent is a declared decision, not a leak.
• Sweep at checkpoints: at phase end and during [[compaction-handoff]], list running agents and stop every finished one before moving on. Compaction wipes the orchestrator's own ledger of open workers, so the sweep must use an enumeration probe (a TaskStop against a nonexistent id lists running teammates), not memory.

Collect results as a file, not as a return message

A worker's final message is not durable transport — long results get truncated silently, and what was cut is gone. This burned three separate times — most recently a panel run that lost three reviewers' output — before it was codified as a gate. The contract: the worker writes its deliverable to a scratchpad/output file whose path the orchestrator fixes in the spawn prompt — collection is a spawn-time decision, because a truncated reply cannot be retrofit into a file after the fact. The final message then carries only the file path plus a short summary. Treat any delegation whose prompt names no output file as not yet dispatched.

V&V separation

The implementer never verifies its own work — the instance that wrote something never judges it. How that plays out per lane:

• Sonnet implemented (반복·단순 구현) → Opus cross-verifies — a different model has different blind spots, which is the cheapest anchoring break available.
• Opus implemented (핵심 구현·테스트) → a fresh Opus instance verifies: a NEW agent with no shared history. A verifier that watched the implementation happen inherits the implementer's mental model and anchors on it — same-session self-review reliably misses the same edge cases the implementation missed. Instance separation is what makes "Opus builds AND Opus verifies" coherent.
• Orchestrator layer on top: the orchestrator hunts 성능·보안 문제점 in shipped features — a second, higher-altitude pass that judges what a diff-level verifier doesn't (product fit, systemic risk). Documents get the same split: Opus authors, the orchestrator reviews via [[multi-persona-review]].

This pairs with, not replaces, deterministic gates (tests, typecheck, CI) — the verifier judges what automation can't: spec fit, missed edge cases, design drift.

Orchestrator handoff (quota exhaustion)

When the top-tier orchestrator's quota runs out mid-project, Opus @ max takes over orchestration. There is no reliable automatic path — documented fallback chains explicitly exclude rate-limit errors, and plan-level auto-switching is undocumented behavior you must not build on. Hand off manually:

1. Run the [[compaction-handoff]] protocol: persist durable facts to memory, take an atomic git snapshot (clean tree + open-PR check), emit the fixed-field resume anchor (current state / verified / what's left / next action).
2. The successor session starts on opus at max (/model opus + /effort max), reads the anchor, and continues as orchestrator under this same policy — setting direction, reviewing, and hunting problems itself (it can author documents directly too, since it already runs at the Opus doc-author tier — no separate delegation needed while it stands in).
3. When the top-tier model becomes available again, hand back the same way.

Anti-patterns

Quick reference

방향성 수립 / 설계·기획 확정 / 작업 분배 / 스펙 리뷰(multi-persona-review) / 기능 개선 / 성능·보안 문제발굴
                               → 오케스트레이터(Fable) 직접
스펙 문서 초안(Fable 브리프 기반) / 핵심 구현 / 테스트 작성·실행(E2E 포함) / 코드 리뷰·검증(V&V)
                               → opus  @ xhigh (또는 max)   — pinned role 또는 Workflow opts
반복·단순 구현(확립 패턴 적용) / 리서치 스윕
                               → sonnet @ high 이상 — 테스트·검증 투입 금지, 산출물은 Opus 교차검증
결정적 변환 (rename·포맷)      → 모델 위임 금지 — sed/grep/스크립트 직접
판단 잔여 0 + 합격을 명령 하나로 판정
                               → sonnet 기본 / 다섯 술어 충족 시 외부 실행기 — 최초 1회 사용자 확인, 산출물은 in-harness 교차검증
도구 부재·인증 만료            → 레인을 내리고 무엇을 못 썼는지 보고 — 대신 설치/로그인 금지, 조용한 제공자 교체 금지
위임 완료                      → 결과 수거와 동시에 TaskStop (SendMessage 재사용 예정 시만 유지 선언)
Fable 소진                     → compaction-handoff → opus @ max 가 오케스트레이터 대행