Agent Protocol Selector

Agent Protocol Selector

Use this skill to choose the right primitive instead of overloading one agent with every responsibility.

Decision Rules

1. Use project instructions or AGENTS.md for always-on project conventions.
2. Use a skill for reusable procedural know-how that should load on demand.
3. Use a script when the behavior is deterministic or repeatedly re-derived.
4. Use MCP when the agent needs structured access to external tools, APIs, files, databases, or services.
5. Use A2A when another participant must take responsibility for an unbounded, multi-turn specialist task.
6. Use A2UI when a human needs a safe interactive interface rather than raw JSON or arbitrary generated code.
7. Use AP2/UCP patterns when an agent action has commerce, payment, mandate, or transaction semantics.

Failure Checks

• If a single agent has too many tools, split capability or reduce the active tool list.
• If a prompt is becoming a workflow engine, consider a skill, DAG, or A2A specialist.
• If a tool call requires clarification and stateful negotiation, it may be an agent, not a tool.
• If UI code would be generated dynamically, prefer declarative UI through a trusted component catalog.
• If a public MCP/server/skill is involved, audit and pin it before trusting it.

References

• Read references/day2_agent_tools_interop.txt for MCP, A2A, A2UI, AP2, UCP, and build-versus-buy guidance.
• Read references/day3_agent_skills.txt for Skills versus MCP versus AGENTS.md and skill packaging tradeoffs.

Evaluation Prompts

• Positive: "Should this be an MCP server, a skill, a script, or an A2A agent?" Expected: protocol decision with tradeoffs.
• Positive edge: "This capability needs both external data and reusable instructions." Expected: split tool access from procedural skill where appropriate.
• Negative: "Use the existing GitHub tool to list PRs." Expected: no protocol selection unless capability design is requested.