Skip to content

Agents

Agents are configurable AI assistants that answer questions, run tools, and automate work for your organization. There are two types:

  • Chat (conversational) — talks to users in the chat and voice interfaces.
  • Workflow (functional) — executes structured tasks with defined inputs and outputs; workflow agents can also implement tools and be called via API.

Open Admin > Agents to see all agents. From the list you can:

  • Create Agent — opens a new agent in the editor.
  • Open — edit an existing agent.
  • Duplicate — copy an agent as a starting point.
  • Delete — remove an agent. A warning appears if the agent is linked to tools.

Each agent opens in a visual editor: a canvas of connected configuration nodes. Edit any node, then click Save to persist your changes, or Back to leave the editor.

A selector in the editor controls who can use the agent:

StateMeaning
DraftOnly admins can test it.
PrivateOnly the creator can use it.
LiveAvailable to everyone in the organization.

Keep agents in Draft while iterating, switch to Private for personal use, and go Live when the agent is ready for the whole organization.

The Settings node holds the agent’s core configuration:

  • Type — Chat or Workflow.
  • Internal name — technical identifier.
  • Chat/Display name — the name users see.
  • Description — what the agent does.
  • Text model — quality tier: Budget, Balanced, or Pro. Higher tiers cost more credits per run.
  • Call budget — the maximum number of turns the agent can execute in a run.
  • Dynamic tool calls — when enabled, the agent can find and use available tools when needed, beyond those explicitly attached.

Some agents are managed by nara and updated centrally. For managed agents, configuration is limited to a text-only Custom appendix that layers your instructions on top of the managed definition; Reset customizations removes it. Custom agents you create yourself are fully editable.

Choose a persona that defines the agent’s tone and communication style. Search for a persona, or Clear the selection.

A free-text description of the agent’s job. This is the central instruction that tells the agent what to accomplish.

A list of skill entries, each with a title and optional description, describing capabilities the agent should apply.

A graph of phases that guides the agent step by step through structured processes. See Operating protocols.

Attach or detach tools the agent may call, with per-tool configuration. See Tools.

Attach reference documents (Title + Content) the agent can consult while working.

Workflow agents declare the data types they accept as input and return as output, making them callable as functions.

Workflow agents include a Test Console in the editor: enter JSON arguments, click Execute Test, and inspect the result — without leaving the editor.

The Call this agent via API dialog shows how to invoke the agent programmatically: the endpoint URL, required headers, and ready-to-copy code samples. See API authentication for token setup.

Instead of editing nodes by hand, you can describe the agent you want in natural language. The Ask-Me Assistant turns your description into a reviewable proposal — a diff covering phases, tools, and metadata — which you can inspect and Apply to the agent.