Build & Evals

The Agent Designer

The Agent Designer is the visual way to author an agent. It reads and writes the same Agent.md the rest of the platform uses — the file stays the source of truth — so anything you compose here is a plain, versioned definition, never a hidden format.

Composing an agent in the Agent Designer.
Composing an agent in the Agent Designer.

Three ways to author, one artifact

  1. The editor — write Agent.md directly, with live validation surfacing schema, reference and policy errors as you type.
  2. The designer — compose the same definition on a canvas: pick the model and memory, set autonomy and the turn cap, add tools, skills, knowledge and MCP servers as nodes, and see what may invoke the agent. Every change is written back to Agent.md, and the editor and the canvas can be used in turn on the same draft.
  3. The Architect — a built-in meta-agent that interviews you, drafts a valid definition, and suggests reusing an existing agent over building a near-duplicate. It writes to a draft buffer and cannot save on its own: a draft becomes an agent only when you accept it.

What the designer shows

The canvas makes the parts of an agent legible at a glance: the model and autonomy level, the tools it declares, the skills it composes, the knowledge it can read, and any MCP servers it reaches. Each is validated — a skill reference that is not visible to you, a tool that is not allowed by policy, or a model outside the allowlist is flagged before you ever save.

The Agent Designer canvas for the built-in Market Watcher agent: the model and memory nodes above, the agent card with autonomy and turn cap, its six tools fanned out below, and the surfaces that invoke it on the left.
The built-in Market Watcher on the canvas: model (default tier) and memory (none) above the card, autonomy supervised with a 20-turn cap, the six tools it declares — market_data, render_chart, web_search, web_fetch, calculate, current_time — and, on the left, what may invoke it (chat, the orchestrator; automations and A2A are off).
Validation is the same pipeline everywhere The designer, the raw editor, and the API all run one validation pipeline, so a definition that is green here is green at create time — there is no second set of rules to be surprised by.

From draft to running

Saving creates the agent in your personal workspace at version 1, where a security scan runs automatically. From there you test and certify it in the Playground, and promote it up through team and org rings when it is ready — see Build & promote.

← Agents Tools →