Glossary

Same word, different products

Four words exist in every product and mean something different in each. Where these docs use them without a product name, they mean the Mindset version.

Terms

Acceptance test

A situation you put an agent in, plus what must be true about how it behaved. Each test runs several times and passes on a threshold. By default a model-judged check runs five times, and a code check once. Also called a BAC (behavioral acceptance criterion). You write, run and retire them in Govern → Optimize.

Example: "Given an invoice with a $40 difference, the agent calls the invoice-comparison function."

Not: A gate the platform enforces. Activating a version doesn't run the tests. Once accepted, a test can't be edited. You amend it (which replaces it) or retire it with a reason.

Agent

The thing that does the work. It has a prompt, a model, the resources it may reach, and optionally a script. People talk to it in the Hub, and schedules, triggers, other agents or your own systems can start it.

Example: The invoice exceptions agent compares invoices to purchase orders and registers supplier queries.

Not: Something that needs a script. Most agents are instructions plus the right operations.

Agent Builder

The assistant docked on the left of an agent's tabs. Describe a change in words and it edits that agent.

Not: Orca. Orca provisions things you don't have yet.

Agent Management Studio (AMS)

The Mindset app where you build, connect, run and govern agents. Its main areas are Agents, Orca, Connections, Functions, Widgets, the Govern pages (Monitor, Log, Optimize, Audit), the Hub and Settings.

Availability

Whether an agent can be reached, and from where. Set on the agent's Versions & Availability tab: one Available switch, plus six places it can be used (The Hub, Embedded in your own product, Other agents, Schedules, Triggers, Outside systems). It applies as soon as you save and isn't versioned.

Connection

A link to one outside system, with the credential for it. The credential stays with Mindset and never reaches the agent or the model. The outside system sees the connection's shared credential, except with StackOne, which acts as the end user. There are seven types: MCP server, Knowledge base, Postgres database, Google Sheet, HTTP API, StackOne, and AI model (LLM).

Example: The finance system, and the supplier contracts knowledge base.

Not: What you give an agent. You give it operations.

Environment

A named partition inside your org. Agents, connections, triggers and schedules each live in one environment. Every org starts with one called Sandbox. Org admins create more and can rename them. The slug in its address never changes, and an environment can't be deleted.

Not: A copy of your work. Nothing moves between environments on its own.

Experiment

A comparison run of an agent's acceptance tests with one thing changed, such as the model or part of the prompt. It reports, per test, whether the change improved, regressed or made no detectable difference, and what each arm cost. It changes nothing on the live agent.

Function

A fixed list of steps that gives the same answer for the same input. An agent calls it as a tool. A step calls an operation, calls a model, or works on data it already has. Functions keep versions and must pass validation before they publish.

Example: The invoice-comparison function fetches the PO, extracts both sets of lines, matches them and returns the differences and their total.

Not: Something you schedule. Schedules and triggers start agents.

Gate

The condition a script phase must meet before the run moves on. A machine-checked gate tests what was collected, with no model involved. A judged gate asks a model to decide, and records its reasoning. A gate can combine both.

Example: Phase one passes when both the invoice and its purchase order have been retrieved.

Govern

The part of AMS where you see what happened and improve it: Monitor (cost, performance, resources and topology), Log (every session, in order), Optimize (acceptance tests and experiments) and Audit.

Hub

Where people in your org talk to agents. Every agent with The Hub ticked is visible to every member, in that environment. Members can also reach Hub agents from their own AI client over MCP.

Not: Where you build. That's the Agents area.

Model key

Your org's own credential for a model provider (Anthropic, OpenAI or AWS Bedrock). Set by an org admin in Settings → Model keys. See Use your own model provider keys on AWS.

Operation

One named thing an agent can do on a connection, with a fixed input and output, marked read or write. Operations are discovered from the system or written on the connection. An operation must be enabled before any agent can call it.

Example: "Get the purchase order for an invoice number" (read) and "register a supplier query" (write).

Operation approval

A write operation needs consent before anything can call it. Consent is given once per operation, not per call. The org setting Turn on connections that change other systems automatically is on by default, so new write operations are enabled without a person. When it's off, a person approves each new write operation on the connection's Operations tab.

Not: Something an agent can do. No agent tool can approve or withdraw an operation.

Orca

The agent that provisions what you don't have yet. Describe what you want, and it looks at what already exists, proposes a plan and stops for your approval. Specialist builders then create each piece.

Not: Where you edit something that exists. That's the Agent Builder.

Org

Your Mindset organization, and the isolation boundary. Members, settings, model keys and audit sit at org level. Its region comes from the deployment it lives on and can't change.

Personal agent

An agent a member builds for themselves in the Hub. Off by default, and turned on in Settings → Defaults. A personal agent holds no connections or other resources. It can only hand work to agents already available to its owner, and it is never shared with anyone else.

Phase

One step of a script, defined by what must be true before the run moves on. A phase says what the agent is told, what it Can use, what it Collects before advancing, and the gate to pass. Routes only go forward.

Example: Gather, Compare, Assess, Prepare.

Run

One execution of an agent, from whatever started it to however it ended. Every run is recorded: what started it, the phases, the operations called, what came back, and any runs it handed to other agents. A run can park while a script waits for a person, and resume when the answer comes.

Script

The ordered phases an agent works through, for the stretch of work that must wait or happen in order. A script has no versions of its own. The agent's version records which script version it uses.

System prompt

What the agent is and how it behaves, on every turn. In AMS it has five parts: Purpose, Personality, Background, Policy rules and Output formatting.

Not: The place for a rule that needs a definition. "A material discrepancy" belongs in a function as a threshold.

Version

A saved snapshot of an agent's configuration: prompt, model, resources, script and schedule. Saving creates a version and activates nothing. Activating a version makes it live from each person's next turn, and rolling back is activating an earlier one.

Widget

Something an agent shows a person in a conversation instead of plain text. Widgets keep versions, must pass validation (it will render) before they publish, and roll back by re-activating an earlier version.

Write

An operation that changes something in an outside system. See operation approval.

Example: Registering a supplier query.