Quickstart
Build, deploy, and run your first AI agent in about ten minutes.
This walkthrough takes you end to end: connect a model, build an agent, test it, deploy it, run it, and review the execution. We build an agent here because it's the quickest path to a working result — but workflows are an equal way to build and share the exact same test, deploy, and run steps.
Before you start
You'll need to be signed in to Compass with access to an organization and a project. If this is your first time, see the interface tour to get oriented, then come back.
1. Pick your workspace
The top bar shows the organization and project you're working in — click either to switch. Everything you create in this quickstart will live in that project. If you're not sure, the Default project is a fine place to start.
2. Connect a model provider
An agent needs a model to think with, so first add a provider connection.
Open Connections
In Studio's left rail, click Connections.
Add a provider
Click a provider's card — for example Anthropic — then New connection.
Enter your credentials
Give the connection a name (e.g. "Anthropic — Quickstart"), then work down the steps: click Continue past the optional Gateway; paste your API key and Continue; pick the model this connection will run and Continue; then click Done on Spend & cost (leave the budget blank for no limit). Finally, click Create connection.
Saving calls the model once to confirm it works, and the connection shows a Generation badge — it can power an agent. See Providers for the full list and options.
3. Build the agent
Create a new agent
In the left rail, click Agents, then New Agent. This opens the agent builder. Your work auto-saves as a draft as you go.
Name and describe it
Give it a clear name, like "Release Notes Writer", and a short description.
Write its instructions
In Instructions, tell the agent what it does and how to behave — this is its system prompt. For example:
You are a concise technical writer. Given a list of merged changes, write friendly release notes grouped into Features, Fixes, and Improvements.
Choose its model
Under Model, select the provider connection you created in step 2.
That's a complete agent. (Optionally, use the Tools card to let it call an integration, a workflow, or another agent — see Giving agents tools.)
4. Test it
Click Test to open the test panel. Type a message — for example, paste a few bullet points of changes — and send it. The response streams back live, with any tool calls, the tokens and cost, and a Trace button for the full record.
The test conversation is throwaway, so experiment freely — though each turn does run for real, and shows up in Observability. Tweak the instructions and test again until you're happy.
5. Deploy it
When it's ready, click Deploy. A dialog confirms the details and runs a quick validation:
- Fix any errors it flags (for example, a tool without a connection). Warnings won't block you.
- The dialog shows the version you're about to publish (your first deploy is v1).
Confirm to freeze this draft into version v1 and publish it as live. Your agent can now be run by your team and other agents.
6. Run it
The quickest way to run a deployed agent is Compass Chat:
Open Compass Chat
Click Chat in the top bar. A new chat opens with What should we work on?
Add your agent
Click Tools in the composer and tick your agent under Agents. It's set to Live · v1 — the version you just deployed. (The list shows the project named in Bills to Project under the composer; if that's a different project, pick a model from yours first.)
Ask it to work
Send a message asking chat to have your agent do its job. Chat calls your agent, and the call appears as a row you can expand to see what went in and came out.
Other ways to run
You can also run on a schedule (see Triggers & schedules) or from your own code through the API (see API, CLI & SDK). A one-off "Run now" button in the UI is on the way.
7. See what happened
Click Trace under the reply to open the full record right there. Or switch to Console — click Studio, the first crumb after the Compass logo, and choose Console — and open Observability: chat turns are listed as Compass Chat Agent, with your agent's run inside. Click any execution to open its trace — the step-by-step record of everything that happened, including each tool call with its inputs and outputs.