Use an Agent Skill for workflow guidance, then connect the agent through MCP, an SDK, or the REST API.
Choose an interface
- MCP is the fastest path for assistants and coding agents that support remote MCP servers.
- SDKs provide typed Search and Act operations in TypeScript and Python.
- REST works from any trusted server environment.
All three interfaces use the same capability IDs, revisions, Action states, and approval rules.
Keep authority explicit
Install instructions and tool descriptions do not grant access. The client still needs an authenticated Darwin connection or a properly scoped API credential. Keep user approval, provider authentication, and payment authorization separate.
Model the workflow
Store identifiers in application state rather than asking the model to recover them from prose. Give the AI the smallest operation-specific schema it needs for the current turn.
Define a failure policy
- Retry only transient transport failures.
- Reuse the same
requestId only for an identical mutation retry.
- Stop autonomous progress when Darwin requires a person.
- Never turn a nonterminal Action into a successful agent answer.
- Recover from a stored
actionId after a restart or timeout.
Last modified on September 21, 2026