Create an agent

Creating an agent is three decisions: what it is called, what it may do, and which model drives it.

Create it

In your app, open Agents and choose New agent. Give it a name and a description of its job. The description is guidance for the model, not a permission — it shapes behaviour, it does not grant anything.

Give it a role

An agent with no role can do nothing. Assign it a role exactly as you would a person, and the policies bound to that role become its tools.

Start narrower than feels useful. It is easy to add a policy once you can see what the agent tried to do and could not; it is harder to notice something it should never have had.

Two agents in one app, each showing its assigned role and status — Two agents over the same data. One may update records and move money but must ask; the other cannot, because those actions are not in its grant at all.

Connect a model

Bind a model to the agent under Connect an LLM. You supply your own OpenAI or Anthropic key — the model provider bills you directly, and the key is stored encrypted.

Connected AI providers listed under Integrations, showing OpenAI and Anthropic connections — Connect a provider once for the organization, then bind it to any agent.

Decide what needs approval

Set the agent's approval posture. Auto-approval is a set of specific triggers rather than a single switch: anything outside the set halts and waits for a person. See Approvals & halts.

Try it

Talk to the agent from the console first. The run timeline shows each step, its arguments, and its outcome — including anything it attempted and was refused, which is the fastest way to find a missing policy.

Checking what it can actually do

Do not infer an agent's reach from its description. Ask it:

curl "$AUTHORIC_URL/$APP_UID/agent/tools" -H "Authorization: Bearer $AGENT_KEY"

That list is the truth — derived from policy at the moment you asked.

Approvals & halts

The boundary and how to resume.

API keys

Let something outside the console drive it.