One Tool, Multiple Modes
Agentforce Builder lets you draft an agent conversationally with AI guidance, refine in a document-like editor with autocomplete, adjust visually in a low-code canvas, or drop to a pro-code script view. Same underlying metadata; different surfaces for different stages of the build. Toggle between modes without losing state — every change writes through to the canonical metadata XML, so a topic added in conversational mode is editable in canvas mode and inspectable in script mode within the same session.
Why It Matters
The previous Agent Builder forced switching between separate views for topics, actions, and prompts, with manual save and reload between each. Builder unifies the loop: draft, test, refine, test. Iteration time drops meaningfully — Salesforce’s internal benchmark shows median time-to-first-eval down from 4.5 hours to 1.2 hours for a 5-Action agent. The other gain is reviewability: pro-code mode emits diffable YAML so source control finally works the way developers expect.
Mode Best for Output
---- -------- ------
Conversational Empty-state scaffolding Topics + Actions skeleton
Document Editing instructions and prompts Inline annotations
Canvas Visualizing topic/action graph Same metadata, visual layer
Script Code review, source control, CI YAML metadataConversational Draft
You describe the agent in natural language — “handle password reset requests, look up user in Identity, escalate if locked, log every reset to a custom object” — and the builder scaffolds Topics, Actions, Instructions, and a starter eval set. Useful for first drafts; still needs review for production. The scaffold defaults to conservative scopes and includes a placeholder Refusal block that you must explicitly approve before publishing.
Document Mode Power Features
Inline @mentions to reference Actions and Knowledge articles, autocomplete on field API names from the agent’s data scope, and live linting that flags ambiguous instructions (“when the user asks for help” — about what?) before you save. Document mode also surfaces “instruction conflicts” — two Topics or Actions whose triggers overlap — and offers a one-click reconciliation.
Canvas and Script Modes
Canvas renders the Topic/Action graph with live edge weights from the eval set: thicker edges where Atlas more frequently routes traffic, dashed edges where confidence is below 0.4. Click any node to inspect or edit. Script mode shows the underlying YAML/Apex with full IntelliSense, integrated with the Salesforce CLI for sf project deploy start straight from the editor.
Fit
Best for teams building many agents fast or for orgs that need a single tool both admins and developers can collaborate in. Heavy customization still benefits from pro-code paths through VS Code with the Salesforce Extension Pack, but Builder is the default entry point and increasingly the only place where new authoring features land first.
When to Skip Builder
Skip when your agent has more than ~20 Actions and is owned exclusively by a developer team — at that scale, the visual modes get cluttered and direct YAML editing in VS Code is faster. Also skip for agents under strict change-control where every edit must go through PR review.
What to Do This Week
Open Agentforce Builder, draft a single-Topic agent in conversational mode, then switch to script mode and read the generated YAML — the round-trip is the fastest way to learn the metadata model.
