Skip to main content
Building an agent is mostly writing and refining its system prompt, then testing it live and shipping it through versioning.

​
The build surface

Open an agent and you land on the Agent tab. Today it exposes:
Voice, LLM, language, and first-message controls are visible in the Agent tab but not yet functional. They’re intentionally inert placeholders until later releases — don’t rely on them yet.
The rest of the tabs (Workflow, Knowledge Base, Tools, Tests, Widget, …) are scaffolding for upcoming capabilities and are not wired in this release.

​
How edits are saved

Editing the system prompt does not immediately change what serves traffic:
  • Each edit autosaves to a draft. Drafts are per-user and per-branch — two teammates editing the same branch don’t collide.
  • The header shows a Draft badge while you have unpublished changes, and an Errors badge if the configuration is invalid (for example an empty prompt or empty agent name). An invalid config can’t be published and blocks creating a branch from it.
  • Publish turns the current draft into a new immutable version on the branch and clears the draft.
  • Discard changes throws the draft away and snaps the editor back to the published baseline.

​
Test before you ship

Use Preview to open a live session and actually converse with the agent using your current draft. This is the fastest way to tell whether a prompt change did what you intended — voice behavior is hard to judge by reading the prompt alone.

​
Iterate safely with branches

When a change is risky, fork a branch instead of editing Main directly. Branch work is isolated; merge it back when the experiment is conclusive. See Versioning for the full model.

​
Next