dex agent new

Scaffold an AI agent project via an interactive Q&A flow.

Synopsis

dex agent new

Q&A flow

dex agent new asks a series of questions to understand the agent’s purpose, and then generates a project skeleton with a CLAUDE.md, system_prompt.md, and starter code.

Questions asked:

1. Name — short identifier for the agent
2. Description — what the agent does in one sentence
3. Trigger — how the agent is activated (user_request, schedule, event, upstream_system)
4. Success criteria — how you know the agent succeeded
5. Reads — data sources the agent reads from
6. Writes — data sinks or side effects
7. Handoff — whether the agent hands off to a human
8. Autonomous — whether the agent runs without human review
9. Example input — a concrete example of what the agent receives
10. Example output — what a good response looks like
11. Bad output — what a bad response looks like (for guardrails)
12. Deploy target — job, serving_endpoint, or interactive

Generated files

<agent_name>/
├── CLAUDE.md            # Agent instructions for Claude Code
├── system_prompt.md     # System prompt template
├── main.py              # Entry point skeleton
└── README.md            # Setup and usage

MCP integration

After scaffolding, the agent project can be served via the dex MCP server:

dex mcp serve

This exposes the agent’s tools to Claude and other MCP clients.