Model orchestration

Model Orchestration Policy

A fixed role split between model tiers, set by the user. 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

Role	Who	Effort floor	Duties
Orchestrator + Product Manager	Top-tier session model (Fable)	session level	Directs everything. Directly: 서비스 방향성 수립·논의, 기획/스펙문서 리뷰, 기개발 기능 개선, 문제점(성능·보안) 발굴 — see below
Doc author + core builder + V&V	opus	xhigh (or max)	기획/스펙/계획 문서 작성·관리, 핵심 구현(novel logic · architecture-touching · security surface), V&V
Repetitive implementation + E2E	sonnet	high (or above)	반복 구현(pattern-following, boilerplate, example-driven migrations), E2E 테스트 작성·실행, research sweeps
Orchestrator stand-in (quota exhausted)	opus @ max	max	Takes over orchestration via a handoff — see "Orchestrator handoff"

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: repetitive vs core

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.

Route to Sonnet (반복 구현 · E2E) when ALL of these hold; otherwise it's Opus core work:

• The task is repetitive or example-driven — a pattern already exists to follow (an adjacent implementation, a spec with worked examples, a test suite to extend). Acceptance criteria fit in a few lines without "use judgment" clauses.
• It's low-blast-radius: boilerplate, N-th instance of an established shape, E2E test authoring/execution, a research sweep. A wrong answer is cheap to detect and redo.
• It needs no architectural or cross-cutting decisions. Actions carry implicit decisions; a worker making design choices in isolation silently diverges from the system (the classic parallel-agent failure).

Core implementation — novel logic, first-of-its-kind shapes, ambiguity, security surface, multi-file coupling — routes to Opus @ xhigh+, as does all verification duty. When unsure, route up — the cost delta is smaller than a redo.

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.

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.
• Collect the result as a file, not as a return message. A long final message can be dropped in transit, and the failure is silent — it reads as "the worker produced nothing", not "the transport lost it", so you retry the work instead of the delivery. Name a scratchpad path in the spawn prompt, have the worker write there (an early partial, overwritten as it refines), then read the file. This is a spawn-time decision: retrofitting costs a re-request per worker, and one that has already ended may not reproduce what it had. Inline returns are for short answers you can afford to lose.

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 구현/E2E) → Opus @ xhigh verifies (cross-model + fresh context — catches classes of bugs same-model review does not).
• Opus implemented (핵심 구현) → 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.

Verdict vocabulary (fixed). A verifier's report never ends in free prose — it ends with exactly one verdict, PASS / PASS_WITH_NITS / FAIL, and every finding carries one severity label, CRITICAL / HIGH / MEDIUM / LOW (full contract: the verification-loop skill's Verdict Contract). Two consequences for orchestration:

• FAIL closes only on re-verification. Fix → re-verify is one cycle; if the same reviewer will re-verify, keep it alive via SendMessage and declare that intent (worker lifecycle) — otherwise spawn a fresh verifier.
• PASS_WITH_NITS is not silent PASS. Its LOW/MEDIUM findings come back to the orchestrator as recorded follow-ups, not as buried prose — deciding their fate is an orchestrator judgment call, never the verifier's. (The name says nits, but the tier it covers is LOW and MEDIUM.)

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

Anti-pattern	Why it's a violation
Agent(model: "opus") with session at default effort	Inherits high < xhigh floor — use a pinned agent role, Workflow opts, or raise session effort
Delegating 방향성 수립 or a final judgment call to a worker	Direction is the orchestrator's own — a delegation prompt can't carry the shaping context
Accepting an Opus-authored spec/plan without orchestrator review	Author≠reviewer applies to documents too — run [[multi-persona-review]] before accepting
Orchestrator hand-writing full spec/plan drafts itself	Authoring is Opus's lane — brief the direction, delegate the draft, review the result
Sonnet on core work (novel / architectural / security)	Fails the routing test; route up to Opus
Effort below floor "to save tokens"	The floors ARE the policy; surface cost concerns to the user instead
Implementer verifying its own diff	Anchoring — verification needs fresh context, prefer a different model
Two agents writing the same files in parallel	Conflicting implicit decisions; keep writes sequential or worktree-isolated
Spawning an agent for what one direct tool call answers	15× token multiplier for zero value — do trivial work directly
Relying on plan-level auto-fallback for continuity	Undocumented behavior; use the manual handoff protocol
Finished worker left running after its result is consumed	Subagent panes/windows accumulate (iTerm2 등) and idle pings pollute the session — TaskStop as one motion with consuming the result

Quick reference

방향성 수립 / 스펙 리뷰(multi-persona-review) / 기능 개선 / 성능·보안 문제발굴
                               → 오케스트레이터(Fable) 직접
기획·스펙·계획 문서 작성·관리 / 핵심 구현 / V&V
                               → opus  @ xhigh (또는 max)   — pinned role 또는 Workflow opts
V&V 판정                       → PASS | PASS_WITH_NITS | FAIL + CRITICAL~LOW (verification-loop 계약)
반복 구현 / E2E 테스트 / 리서치 스윕
                               → sonnet @ high 이상
결정적 변환 (rename·포맷)      → 모델 위임 금지 — sed/grep/스크립트 직접
위임 완료                      → 결과 수거와 동시에 TaskStop (SendMessage 재사용 예정 시만 유지 선언)
Fable 소진                     → compaction-handoff → opus @ max 가 오케스트레이터 대행