Documentation

Agent Designer

View as Markdown

Agent Designer, also called Describe With AI, creates a complete Agent from a natural language description. Instead of manually configuring every step, variable, loop, and output, you describe the work the Agent should perform and the designer drafts the Agent for review.

How It Works

Designer uses a two-phase flow.

Phase 1: Clarifying Questions

After you describe the Agent, the designer may ask clarifying questions about inputs, outputs, sources, tools, schedules, or review requirements.

Questions may be multiple choice or open-ended. Answer them and send the responses to continue. If your initial description is detailed enough, the designer may proceed directly to generation.

Phase 2: Agent Generation

The designer drafts the Agent, including:

  • Agent variables
  • prompt steps
  • loop steps for collections
  • JSON validation or structured output settings
  • model and reasoning assignments
  • source connection expectations
  • tool usage where appropriate
  • benchmark configuration
  • final output selection

The generated Agent is a starting point. Review it before running it on sensitive content or scheduling recurring execution.

Write Effective Descriptions

Describe three things clearly: inputs, outputs, and processing.

Good input examples:

  • "Input: an uploaded contract PDF."
  • "Inputs: company name, reporting period, and selected Site."
  • "Input: a JSON array of customer names."
  • "Read approved policy folders from the selected Site."

Good output examples:

  • "Output a markdown report with executive summary, findings, risks, and recommended actions."
  • "Extract key terms as JSON with effective date, term length, renewal, liability cap, and termination rights."
  • "Create a spreadsheet with one row per counterparty and columns for risk level, clause reference, and suggested follow-up."

Good processing examples:

  • "For each contract, extract terms, validate the JSON, then summarize risks."
  • "Use Site content only; do not use web search."
  • "Use web search only for current public market data, then cite approved Site documents separately."
  • "Generate a PowerPoint deck after the analysis is complete."

What the Designer Builds

The designer applies Agent design patterns such as:

  • isolated steps with explicit context
  • variable wiring
  • structured outputs
  • loop handling for lists
  • validation steps
  • tool calls where allowed
  • benchmark setup

When imported, the Agent opens in the editor so you can adjust prompts, variables, connections, and policy before running.

Review Before Production

Before scheduling or sharing outputs, review:

  • selected model and reasoning effort
  • permitted source connections
  • controlled tools and egress settings
  • generated file settings
  • benchmark and review gates
  • output template
  • evidence requirements

See Agent Connections, Controlled Tools and Egress, and Benchmarks.