Agents

An agent is a prompt, a model and, optionally, a script. It uses only the connections, functions, widgets and other agents you give it.

An agent talks to people and systems and makes the judgment calls. It has:

Part	Holds
Prompt	Five fields: purpose, personality, background, policy rules and output formatting, up to 2,000 characters each. The prompt applies for the whole run
Model	One of the models your org has a provider key for
Script	Optional. Holds the agent to an order, such as waiting for a person. See Scripts
Resources	The operations, functions, widgets and other agents it may use

Why it matters#

Mindset enforces an agent's resource list whatever the prompt says. Its prompt and model live in Mindset, so changing how it behaves means activating a new version, with no release of your code.

Which one to use#

The work	Put it in
Who the agent is, what it knows, how it behaves	The prompt
A judgment call, such as whether a $340 overage is worth querying	The agent
Work with one right answer, such as comparing an invoice to its purchase order	A function
The same action for every item in a list	A function
A step that must wait, or must not be skipped	A script
A task another agent already does	That agent, added as a resource

A pasted skill becomes prompt text by default.

Orca and the builders#

In the Studio, admins can also build with Orca, the agent in the workspace rail that plans and provisions other agents from a Claude skill, a spec, a description or an MCP server.

Step	What Orca does
Brief	Restates what it read and what it couldn't convert
Inventory	Lists what already exists, to reuse it
Plan	Shows one row per piece, each marked Build it or Use existing, and stops. Nothing is built until you approve
Provision	Builds connections, then functions and widgets, then the agent and its script

Orca creates the agent itself and hands every other new piece to the Connection Builder, Function Builder, Widget Builder or Script Builder. The Agent Builder creates or changes one agent: its prompt fields, model, script, resources, schedules and behavior tests. It can't connect a system or build a function or widget.

Orca matches on names, so a connection named after a project can be missed. Check Connections before you approve.
Long skills are condensed to fit each prompt field. Orca says what it shortened.
Orca tests what it built, and waits for your go-ahead before a test run writes.
Orca activates a new agent's first version. A builder's edits to a live agent wait as a new version for a person to activate.
New agents stay unavailable. A builder ticks only Other agents, Schedules or Triggers when its own wiring needs them. See How agents are started and reached.

Orca and the builders can't:

Approve a write operation. No agent can. See Security and approval.
Run a skill's bundled code. Nothing in Mindset runs a skill's scripts, programs or data files.
Read files on someone's computer.
Drive a website's screens. No connection type logs in and clicks through a site.
Write a script shape the language refuses: a route back to an earlier phase, an action repeated per item, or a model named per phase.

Orca lists what it couldn't convert in its plan and again in its final summary.

To build one from your coding agent, see Build an agent.