Agents
An agent is an LLM-driven worker that carries out a task on your behalf. Every agent is defined by a configuration:
- A prompt that defines its role and behavior.
- A set of MCP tools it's allowed to call (warehouse, dbt, GitHub).
- A set of skills - reusable, markdown-authored instructions.
- The model it runs on.
Agents run against the Anthropic Claude models and stream their work back to the UI in real time, so you see reasoning, tool calls, and file changes as they happen.
Built-in agents
| Agent | Where you use it | What it does |
|---|---|---|
| Build agent | Studio | Creates and modifies dbt models, validates them, and opens PRs. |
| Refactor agents | Refactor | A family of task-specific agents that execute pipeline steps (merge, rename, re-layer, etc.). |
| Optimize agents | Optimize | Scan for issues and generate fix candidates. |
Tools & skills
- Tools are concrete capabilities the agents use: run SQL, introspect the
warehouse, read dbt metadata and lineage, run
dbt compile/test/show, and create or fetch pull requests. - Skills are reusable instruction snippets (markdown) that can be attached to agents to teach them repeatable procedures.
Custom agents
The Agent Builder lets you define your own agents: pick a prompt, choose which tools and skills they can use, and select a model. Custom agents live alongside the built-in ones and can be duplicated and versioned.