dynamic-agents
Installation
SKILL.md
Dynamic Agents
Top-tier output at controlled cost: you and all judgment-heavy work stay on the session model, mechanical work drops to Sonnet, and other model families (GPT via Codex, or Fable) are used only on the nodes where a different or sharper model genuinely earns its cost. Multiple agents buy quality two ways — parallelism for reach, a second model family for a sharper eye — never as a reflex.
1. Decompose
Break the task into subtasks with explicit boundaries: input, required output, what not to touch. Subagents see none of the conversation, so every prompt must be self-contained (paths, constraints, output format). Classify each subtask as judgment-heavy, mechanical, or second-opinion before routing.
2. Route models
| Tier | Use for | How |
|---|---|---|
| Session model (inherit) | design, architecture, synthesis, non-trivial implementation, debugging, final decisions, anything ambiguous | omit model |
| Sonnet | mechanical work with a clear spec: codebase searches, fact gathering, research with a defined question, scripted browser/device QA (explicit steps and pass criteria), low-risk well-specified single-file edits, boilerplate from an existing pattern | model: "sonnet" |
| Fable (escalation) | the one or two hardest reasoning nodes — a thorny design call, a subtle bug, a Claude-family second opinion distinct from Codex — when the session model is not already Fable | model: "fable" |
| GPT-5.6-sol+ via Codex CLI | cross-model code review, second opinion on a diagnosis, getting unstuck, copywriting drafts | subagent_type: "codex:codex-rescue" (Agent tool) or agentType: 'codex:codex-rescue' (Workflow) |