SMELT

Auto Agent

Auto Agent is a special block type that uses the PlannerAgent under the hood to decompose a high-level goal into a sequence of sub-tasks. Instead of manually wiring a DAG of agent blocks, you describe what you want to achieve and Auto Agent figures out the execution plan.

How it works

When a workflow reaches an Auto Agent block, the PlannerAgent analyses the goal prompt and generates a sub-task DAG. Each sub-task is dispatched as Research or Reasoning, the two agent types the Studio palette exposes. Results flow back through the orchestrator and are collected into a single output for downstream blocks.

Configuration

Select an Auto Agent block on the canvas to open its config panel:

  • Goal - describe the desired outcome in plain language. Be specific about what success looks like.
  • Allowed agents - restrict which agent types the planner can dispatch. The Studio palette restricts this to Research or Reasoning; both are enabled by default. Useful for constraining cost on simple goals.
  • Max sub-tasks - upper bound on the number of sub-tasks the planner can create. The slider runs from 1 to 10 and starts at 5. Lower values reduce cost; higher values allow more complex decomposition.
  • Model - the LLM model used for planning. More capable models produce better decompositions for complex goals.

Writing effective goals

Auto Agent performs best with goals that are:

  • Specific - "Research the top 5 competitors in the UK fintech space and summarise their pricing" rather than "Look into competitors"
  • Outcome-focused - describe the desired output, not the steps to get there
  • Bounded - include constraints like timeframe, number of results, or depth of analysis
  • Measurable - clear success criteria let you write a sharper goal and check the output afterwards

Auto Agent vs manual DAG

Use Auto Agent when:

  • The task is exploratory and the optimal agent sequence is unclear
  • You want rapid prototyping before committing to a specific pipeline design
  • The goal is well-defined but the decomposition is complex

Use a manual DAG when:

  • You need the same execution order on every run
  • The pipeline is performance-critical and every credit matters
  • You want full control over which models and tools each step uses

Next: Templates or Builder Guide