An Agent is a configurable AI worker that you can create and deploy to perform specific tasks autonomously. Think of it as a specialized team member you can design, train, and assign to a job. Each agent operates based on a set of instructions and can be equipped with various tools to achieve its objectives. Agents are the core of the Ag.dev platform. They are designed to be “headless,” meaning they execute tasks in the background without requiring a real-time conversational interface. You define their purpose, and they work independently to deliver results.

​
The Anatomy of an Agent

An agent is defined by several key properties that control its behavior and capabilities:
  • Goal Prompt: This is the most important instruction, defining the agent’s primary objective. A clear, specific goal is crucial for success.
  • Tools: An agent can be equipped with a set of Tools to perform actions like searching the web, browsing websites, or interacting with external services via MCP.
  • Input Schema: This is a JSON schema that defines the structure of the data an agent expects to receive for a run. Learn more about Inputs.
  • Output Configuration: You can configure an agent to return its result as a freeform document or as a structured json object that conforms to a specific schema.
  • Model Stack: You can select the underlying AI model that the agent will use, allowing you to choose the best fit for your task’s complexity and cost.
  • Publication Status: Agents display status indicators in the Console to show their current state - Unpublished (no published version yet), Published (live version ready for production use), or Unpublished Changes (published version with newer draft changes). Learn more about versioning.

​
The Agent Lifecycle

All agents are versioned by default. All configuration changes are stored as immutable revisions, allowing you to safely experiment, rollback, and maintain consistent production deployments.
  1. Creation: You create an agent through the Ag.dev Console or API, defining its anatomy. This creates the first revision.
  2. Configuration: Make changes to your agent’s settings. Each change creates a new revision that can be tested before publishing.
  3. Publishing: Promote a tested revision to production, making it the default for new runs.
  4. Execution: Once published, the agent is ready to be run. Each execution is called a Run and uses the published revision by default.
  5. Monitoring & Refinement: Monitor performance through run results. Create new revisions to improve effectiveness, test them, and publish when ready.
By creating and combining different agents, you can build powerful, automated workflows to handle complex tasks at scale.