agent-model-selection（Agent 模型选择）

Agent Model Selection Skill

Purpose

Provides data-driven guidance for selecting the most appropriate language model when creating or modifying agent definitions.

When to Use

• When creating a new agent and need to assign a model
• When modifying an existing agent's model assignment
• When troubleshooting agent performance issues related to model capabilities
• When optimizing costs across the agent ecosystem

Reference Data

Always consult docs/ai-model-reference.md for:

• Current performance benchmarks by category (Coding, Reasoning, Language, Instruction Following, etc.)
• Model availability in GitHub Copilot Pro
• Premium request multipliers (cost)
• Recommended model assignments by agent type
• Task-based guidance and tutorials (via external links with descriptions)

This reference is updated periodically with latest benchmark data.

Critical Learnings

1. Use task-specific benchmarks, not overall scores

- Different models excel at different tasks - Example: GPT-5.2-Codex excels in Coding while Claude Sonnet 4.5 is better for Language (76.00)

1. Claude Sonnet 4.5 has poor Instruction Following (score: 23.52)

- Unsuitable for agents that follow templates (Task Planner, Quality Engineer) - Use Gemini models instead for structured output (scores: 65-75)

1. Gemini 3 Flash offers best value for many tasks

- 0.33x premium multiplier (cost-effective) - Strong Instruction Following (74.86) - Good Language performance (84.56) - Ideal for: Task Planner, Release Manager, high-frequency agents

1. GPT-5.2-Codex is the latest coding model

- Latest generation Codex model (improved over 5.1 Codex Max) - Specialized for agentic coding tasks - Primary choice for Developer agent - Also solid for Code Reviewer

1. Always verify model availability

- Check against official GitHub Copilot documentation - Model names must match exactly (case-sensitive) - Include "(Preview)" suffix for preview models (e.g., "Gemini 3 Pro (Preview)")

1. ⚠️ Coding agents must NOT have model: in frontmatter

- The model: property is only valid for VS Code agents (files without -coding-agent suffix) - On GitHub.com, the model: property in *-coding-agent.agent.md files causes a hard CAPIError: 400 The requested model is not supported error (confirmed by experiment) - The GitHub docs say this property is "ignored" but in practice it prevents the agent from running - Apply model selection only to the corresponding VS Code agent file (e.g., developer.agent.md)

Model Selection Process

When selecting or changing a model:

1. Identify the agent's primary task categories (from ai-model-reference.md)

- Coding, Reasoning, Language, Instruction Following, etc.

1. Check category-specific performance

- Look up relevant benchmarks in ai-model-reference.md - Compare top 3-5 performers in that category

1. Consider cost vs frequency

- High-frequency agents → favor lower multipliers (0.33x, 0x) - Critical accuracy agents → favor best performer regardless of cost

1. Verify availability

- Confirm model is listed in "Available Models" section - Check it's available for VS Code (required)

1. Document your reasoning

- Include benchmark scores in proposal - Explain trade-offs made

Example Model Selection

Scenario: Selecting model for Quality Engineer agent

1. Primary tasks: Define test plans following specific template format
2. Key categories: Instruction Following (critical), Reasoning (important)
3. Benchmark lookup (from ai-model-reference.md):

- Gemini 3 Flash: Instruction Following 74.86, 0.33x cost ✅ - Gemini 3 Pro: Instruction Following 65.85, 1x cost ✅ - Claude Sonnet 4.5: Instruction Following 23.52 ❌ (disqualified)

1. Decision: Gemini 3 Pro (balance of performance and cost)
2. Rationale: Strong instruction following (65.85), reasonable cost (1x), good for template-based work

When to Update Model Assignments

Reassess models when:

• New benchmark data shows significant performance changes
• Agent is underperforming its tasks consistently
• New models are released with better performance
• Cost optimization is needed
• ai-model-reference.md is updated with new data

Key Principles

• Task-specific benchmarks matter more than overall scores
• Balance cost with performance based on agent frequency and criticality
• Always verify availability against official documentation
• Document your rationale for model selection decisions