Core concepts
The vocabulary of Compass — every core idea, what it means, and how the pieces fit together.
This page defines the building blocks of Compass and how they relate. Read it once and the rest of the docs will click into place. Each concept links to its full guide.
Where your work lives
Organization
An organization is your company's tenancy boundary in Compass. Everything — projects, connections, agents, workflows, runs — belongs to exactly one organization, and members of one organization never see another's data.
Organizations are provisioned by Sage and managed through your identity provider; you don't create them inside the app. You can belong to more than one and switch between them from the workspace switcher.
Project
A project is a container inside an organization that holds a related set of
connections, agents, workflows, and deployments. Projects let you separate work —
for example staging from production, or one client from another.
Every organization starts with a Default project, and you can create more at any time. The workspace switcher always shows which project you're working in.
→ Projects
What you connect
Connection
A connection is a saved, reusable link to an external service, with its credentials and settings stored securely. You create a connection once and reference it from any agent or workflow in the project. There are three kinds:
- Provider — an LLM provider such as Anthropic, OpenAI, or LM Studio. This is the "brain" an agent thinks with.
- Integration — a third-party service or data source, such as Postgres or any API you describe with an OpenAPI spec. This is what your AI can act on.
- Gateway — an optional proxy (such as Apigee) that provider calls are routed through for monitoring and policy.
What you build
Compass has two first-class ways to build, and they're peers — you can use either, or combine them.
Agent
An agent is a single AI worker. You give it written instructions, choose a model (a provider connection), and attach the tools it's allowed to use. At runtime it reads its input, reasons, calls tools as needed, and produces a result — either free-form text or a structured object you define.
Workflow
A workflow is a graph of steps you assemble visually on the canvas. Each step is a node; you wire nodes together so the output of one feeds the input of the next. Workflows are the right tool when you want explicit, multi-step logic — fetch data, run an agent, transform the result, call an API — laid out as a diagram.
Node
A node is a single step on the canvas. The main node types are:
- Input — the workflow's entry point and the shape of its input.
- Output — the workflow's result, assembled from upstream steps.
- Agent — an AI step (the same agent concept, embedded in a workflow).
- Prompt — renders a text template with variables from earlier steps.
- Workflow — runs another workflow as a step.
- Integration — calls a connected service.
Tool
A tool is a capability you grant an agent so it can act, not just talk. A tool can be an integration action (e.g. "query the customers table"), an entire workflow, or even another agent. The agent decides when to call its tools while it works; every call is recorded in the run's trace.
How you run it
Version
When you deploy, Compass freezes the current draft into an immutable, numbered version (v1, v2, …). Versions never change, so a run can always be traced back to the exact definition that produced it.
Deployment
A deployment is the live binding of an agent or workflow to one version, plus any schedules attached to it. Think of it as a movable pointer: you promote a version to make it live, and you can roll back to an earlier version at any time. Only deployed work can be run by your team and systems.
Trigger
A trigger is what starts a run. Today you can run from Compass Chat, on a schedule (a cron-style timetable), or through the API. Webhook and event triggers are on the way.
Compass Chat
Compass Chat is a built-in conversational interface for running your work. You pick which agents, workflows, and integrations the chat can use, then talk to it — responses stream in real time, and you can open the trace for any reply.
What you observe
Run
A run is one execution of an agent or workflow. It records the input, the output (or error), how long it took, how many tokens it used, what triggered it, and which version executed. Runs are the heart of Observability.
Trace
A trace is the step-by-step record inside a run: every node that executed, every tool the agent called, the inputs and outputs at each step, and any sub-runs. While a run is in progress you can watch the trace stream live; after it finishes you can explore the whole tree. The trace is how you understand and debug what your AI actually did.