Skip to main content
An agent is a model loop the platform runs. You give it instructions, capabilities, and a routing strategy; the platform handles the turn loop, the conversation state, the autonomy gate, and the audit trail.

​
Agent or gateway?

Both bind the same capabilities. Choose an agent when you want the reasoning run for you; choose a gateway when the client already has a model and just needs tools.

​
Where an agent’s settings live

The Portal opens each agent on a set of settings views:

​
Running an agent

  • Playground — the Portal’s chat surface. Approvals render inline; ?debug=1 on the URL reveals the assembled system prompt and the frozen tool and skill snapshots for the run.
  • Channels — Telegram, Slack, or SMS. See Channels.
  • A2A — POST https://agent-{handle}.noorle.com speaks A2A JSON-RPC 1.0. A public agent card is served at /.well-known/agent-card.json.
  • AG-UI — POST https://agent-{handle}.noorle.com/agui returns an AG-UI event stream.

​
Composition patterns

  • Research — Web Search, Browser, and a knowledge base for internal context.
  • Support — Knowledge Retrieval over your docs, a connector for the systems of record, a Slack channel, and delegation to a specialist agent.
  • Analysis — Files plus Code Runner for anything the standard library covers; Sandbox when a dependency is needed.
Bind the fewest capabilities the job needs. Every bound tool is context the model spends on every turn, and at 30 tools Adaptive presentation switches the surface to progressive discovery.

​
Next