Agents

The agent resource — what it is, its fields, and the two classes agents fall into.

Agents

An agent is the unit that resolves customer work. It carries a type, a description, an activation state, and — for voice agents — a voice model.

An agent on its own does very little. What it’s allowed to do lives in Settings, Rules, and Functions. Create the agent, then constrain it.

Shape

Field	Type	Notes
agent_name	string	Display name
agent_type	string	Which class of agent — see below
description	string	What this agent is for
activated	string	Whether the agent is live
org_id	string	Owning organisation
voice_model	string	Voice model, for voice agents
voice_model_id	string	Its identifier
created_at / updated_at	string	Timestamps

Endpoints

Method	Path
GET	/v1/agents	List
GET	/v1/agents/{id}	Get
POST	/v1/agents	Create
PUT	/v1/agents/{id}	Update
DELETE	/v1/agents/{id}	Delete

Example

curl https://api.stateset.com/v1/agents \
  -H "Authorization: Bearer $STATESET_API_KEY"

curl -X POST https://api.stateset.com/v1/agents \
  -H "Authorization: Bearer $STATESET_API_KEY" \
  -H 'content-type: application/json' \
  -d '{
    "agent_name": "Returns Specialist",
    "agent_type": "automation",
    "description": "Handles RMA creation, receipt, and refunds under $100.",
    "activated": "false"
  }'

Create with activated: "false", configure settings and rules, then activate. An agent that goes live before its allowed_intents are set will act on intents you never approved.

Two classes of agent

The distinction matters because it determines what you configure.

Automation agents

The workhorses. They execute: trigger functions, integrate with systems, and complete repetitive work — order management, subscription renewals, inventory checks, returns processing.

Configure these carefully: they take actions with consequences. What matters is Functions (what they can call), Rules (what’s forbidden), and the review gate (what needs a human).

Analysis agents

These synthesise: reasoning over rich data to surface trends, customer-behaviour patterns, and supply-chain optimisations. ReasonCX is one of these — it emits its reasoning alongside its conclusions so a finding can be traced back to the tickets that produced it.

Configure these for retrieval quality rather than for safety: Embeddings, Memories, and Examples.

The split is about consequence, not capability. An analysis agent that’s wrong produces a bad recommendation you can discard. An automation agent that’s wrong has already issued the refund — which is why the safety machinery is aimed at that class.

Voice agents

An agent with a voice_model answers the phone. The telephony, streaming, and supervision layer is StateSet Voice — including live monitor, whisper, barge, and escalate on an in-flight call.

Related

• Settings — the 22 fields that bound agent behaviour
• Rules — named policy constraints
• Set up ResponseCX — the ordered path to a live agent
• Public API — provision agents programmatically