Notre avis

Ce skill définit un système d'escalade basé sur la confiance entre modèles (Haiku, Sonnet, Opus) pour les agents, avec des règles strictes de progression par paliers et des signaux de déclenchement.

Points forts

Hiérarchie claire des niveaux d'agent avec des rôles spécifiques
Signaux d'escalade précis et actionnables
Règle de non-saut de palier pour éviter les gaspillages
Protocole de diagnostic avant escalade

Limites

Nécessite une configuration manuelle des agents
Ne couvre pas les cas d'erreur d'infrastructure
Peut ajouter de la latence pour des tâches complexes

Quand l'utiliser

Lorsque vous gérez un pipeline d'agents de différents niveaux de capacité et que vous voulez optimiser le coût et la fiabilité.

Quand l'éviter

Pour des tâches simples ou uniques qui ne justifient pas une infrastructure multi-agents.

Exemples

Escalate from Haiku to Sonnet on test failure

I am using a Haiku agent and it failed to fix the test failures in this task. Please escalate to Sonnet with the original prompt and the reason: 'Code introduces new test failures that Haiku could not resolve in one retry.'

Configure escalation for a new project

Set up model escalation for my project: use Haiku for documentation, Sonnet for coding, and Opus for architecture. Never skip tiers and log escalations to claude-mem.

name: escalation description: "Confidence-based model tier escalation for agent tasks. Defines when to escalate from Haiku to Sonnet to Opus, escalation signals, and the never-skip-tiers rule. Use when an agent fails or underperforms."

Model Escalation

Escalation Tiers

Haiku (junior-coder, docs, explainer, optimizer)
  ↓ on failure
Sonnet (coder, tech-lead, reviewer, qa, architect, explorer)
  ↓ on failure
Opus (senior-coder, planner, red-teamer)

Never Skip Tiers

ALWAYS try the next tier first before jumping to Opus
Haiku → Sonnet → Opus (never Haiku → Opus directly)
Each tier gets ONE attempt before escalation
Exception: if the task is known to require Opus (architecture, cross-cutting), start there

Escalation Signals

Escalate to the next tier when:

Test failures: Agent's code fails tests and agent can't fix in one retry
Clippy warnings: Agent introduces warnings it can't resolve
Wrong output schema: Agent produces output that doesn't match expected format
Tech-lead BLOCKED twice: Tech-lead rejects the same agent's work twice on the same task
Task timeout: Agent exceeds expected completion time by 2x
Repeated hook blocks: Agent triggers enforcement hooks 3+ times on the same operation

Do NOT Escalate When

Task is simply large (break it down instead)
Agent needs more context (add skills/instructions instead)
Infrastructure error (retry, don't escalate)
First-time failure on a reasonable task (retry once at same tier)

Escalation Protocol

Diagnose the failure (see capability-diagnostic skill)
Determine if it's a model capability issue (Step 3 of diagnostic)
If yes: re-assign task to next tier agent with original prompt + "Previous attempt failed because: [reason]"
If no: fix the actual issue (context, tools, spec) and retry at same tier

Tracking

Log escalations to claude-mem: "Escalated [task type] from [tier] to [tier] because [reason]"
After 3+ escalations of the same task type → update agent-routing defaults
The optimizer agent reviews escalation patterns periodically

Escalade de modèles par confiance

Recommandé pour