Saltar al contenido
NeoXamAgents

How to declare an agent with a descriptor

From a YAML descriptor to a served agent: the Git-governed path every agent follows.

An agent is fully defined by its descriptor — YAML front matter plus a Markdown system prompt — covering model, tools, skills, input and output schemas, and constraints. No agent without a descriptor: nothing outside it can affect runtime behavior.

From descriptor to production

  1. Start from a template (blank task agent, blank conversational agent, or a product example) — or describe the agent in natural language and let the Quickstart wizard generate a valid descriptor (Beta).
  2. Open a pull request on the Catalog repository — the Git-backed source of truth for agents, tools, skills, environments, and eval datasets.
  3. CI validates the descriptor: required fields, resolvable tool, skill, and environment references, semantic versioning, and mode and memory coherence.
  4. The eval gate runs the agent's eval dataset and compares the result to its baseline score.
  5. On merge, the runtime Catalog serves the new version within minutes — no platform redeploy.

Beta

The Quickstart wizard and natural-language descriptor generation are rolling out with the V1 Beta. The descriptor-as-code Git flow is available today.

Versioning

Descriptors are semantically versioned, and every run records the Catalog SHA it executed — so you can always tell which version of an agent produced a result.

Still stuck?

Tell us what you were trying to do and where it failed — screenshots and run IDs help. A human follows up.