Multi Model Orchestration Template

Multi-Model Orchestration Template

Use this skill when one workflow should intentionally use different models for different stages.

Core rule

Do not assume SKILL.md itself switches models. Model selection happens when you delegate work through a mechanism that supports model override, such as a sub-agent/task runner that accepts a model parameter.

When to use multi-model orchestration

Use it when the workflow has clearly different step types, for example:

• cheap classification first, expensive synthesis later
• broad scraping first, deep reasoning later
• repetitive formatting first, careful final writing later
• automation steps on default model, analysis steps on stronger model

Default pattern

Split work into 3 lanes:

1. Main agent lane

• Owns user context
• Decides routing
• Calls tools
• Verifies outputs
• Writes final answer

2. Fast/cheap lane

• Use a lower-cost or default model
• Best for:
  • triage
  • extraction
  • cleanup
  • simple transforms
  • repetitive batch steps

3. Strong reasoning lane

• Use a stronger model such as sonnet
• Best for:
  • planning
  • synthesis
  • ambiguous decisions
  • writing high-stakes output
  • strategy or ranking

Recommended decision framework

Before delegating, classify each step by:

• cost sensitivity: low / medium / high
• reasoning depth: shallow / medium / deep
• risk of mistakes: low / medium / high
• need for user-context continuity: low / high

Then route:

• shallow + low risk → default or cheaper model
• deep reasoning or high stakes → stronger model
• actions touching user context / external side effects → main agent supervises

Reusable orchestration skeleton

Pattern A: Main agent routes to one stronger sub-agent

Use when only one phase needs a better model.

Example:

1. Main agent gathers inputs.
2. Main agent delegates planning/synthesis to sub-agent with sonnet.
3. Main agent executes tools and returns final result.

Pattern B: Fan-out + synthesis

Use when many items need light processing, then one final synthesis step.

Example:

1. Main agent splits items.
2. Fast lane processes each item.
3. Strong lane synthesizes results.
4. Main agent verifies and responds.

Pattern C: Automation + judgment

Use when browser/tool work is deterministic but interpretation is not.

Example:

1. Main agent performs browser automation and data capture.
2. Strong lane analyzes the captured data.
3. Main agent applies the analysis back into tools or final response.

Suggested wording inside a skill

When authoring another skill, use language like this:

• Use the default model for browser automation and stateful page interaction.
• If a reasoning-heavy ranking/synthesis step is needed, spawn a sub-agent with model alias sonnet.
• Keep final user-facing judgment in the main agent unless the delegated task is purely analytical.

Template snippets

Snippet: route by step type

If the step is mostly extraction, formatting, or bulk processing, keep it on the default model.
If the step requires deep comparison, nuanced judgment, or high-quality writing, delegate to a sub-agent using model alias `sonnet`.

Snippet: supervised delegation

Delegate analysis to a stronger model, but keep tool execution, external actions, and final verification in the main agent.

Snippet: batch then synthesize

For N similar inputs, do lightweight extraction first. After collecting structured outputs, use a stronger model once to rank or synthesize.

Example mapping

Example 1: Research workflow

• Search + scrape → default model
• Compare options and rank → sonnet
• Final concise recommendation → main agent

Example 2: Browser automation workflow

• Navigate/click/upload/download → default model
• Interpret messy page results / summarize findings → sonnet
• Confirm final files and report status → main agent

Example 3: Inbox triage workflow

• Message classification → default model
• Sensitive reply drafting → sonnet
• Sending / routing → main agent

Guardrails

• Do not use a stronger model for everything by default.
• Do not delegate actions with external side effects unless supervision is explicit.
• Do not lose track of source-of-truth files, outputs, or user constraints when splitting work.
• Prefer fewer, larger delegations over many tiny ones.
• Record in the parent flow which step used which model if traceability matters.

Minimal authoring checklist

When adapting this template into a real skill, define:

• which steps stay on the main/default model
• which steps should use a stronger model
• what input/output contract each delegated step must follow
• what the main agent must verify before replying