The Agent Builder
Every agent opens into the same builder — a top bar for saving and publishing, and a row of tabs for everything else. This page orients you to the layout; each tab has its own guide.
The top bar
The top bar shows the agent's name and its current state, and holds the two actions that move it between draft and published:
- Save draft — saves your in-progress changes. Disabled until you have unsaved changes.
- Publish — publishes the current draft as the version other people and workflows use. Disabled until a draft exists.
Next to the name, a tag shows where the agent stands:
| Tag | Meaning |
|---|---|
| Editing draft vN | You have unsaved or saved-but-unpublished changes. A second line shows Published version live is vN or Not published yet. |
| Published vN | No open draft — what's published is what runs. |
| Not published | Never published. |
See Publishing an Agent for what publishing does and how draft versions work.
If you navigate away with unsaved changes, Studio asks Leave without saving? before discarding them.
The tabs
| Tab | What it's for | Guide |
|---|---|---|
| Chats | Test the published version in a live conversation | Testing an Agent |
| Configuration | Instructions, run settings, and A2A delegation | Configuring an Agent |
| Capabilities | Connection, tools, MCP sources, skills, prompts, and guardrails | Attaching Capabilities |
| Knowledge base | Files the agent can use as context | Adding a Knowledge Base |
| Memory | What the agent remembers across runs | Configuring Memory |
| Logs | Every reply the agent has given | Reviewing Logs |
| Evals | Automated test runs against saved datasets | Running Evals and Datasets |
| Datasets | Reusable test cases for evals | Running Evals and Datasets |
Where to start
If you just created the agent, go to Configuration and Capabilities first — an agent needs a connection before it can chat, and instructions before it's useful. Once it behaves the way you want, publish it and try it from Chats.