sf-ai-agentforce（SF AI 特工部队）

sf-ai-agentforce: Standard Agentforce Platform Development

Use this skill for the Setup UI / Agent Builder path: declarative topics, Builder-managed actions, GenAiFunction / GenAiPlugin metadata, Prompt Builder templates stored as GenAiPromptTemplate metadata, Models API usage from Apex, and custom Lightning types.

For new code-first agent development, prefer sf-ai-agentscript. If the work produces or edits a .agent file — including Builder Script / Canvas work that results in an authoring bundle — use sf-ai-agentscript.

When This Skill Owns the Task

Use sf-ai-agentforce when the user is:

• maintaining existing Builder-based agents
• working in Setup → Agentforce → Agents
• creating or fixing GenAiFunction, GenAiPlugin, or GenAiPromptTemplate metadata
• wiring Builder topics to Flow / Apex / Prompt Builder actions
• using Models API or LightningTypeBundle in the context of Builder-based agents

Do not use it for:

• .agent files or deterministic FSM design → sf-ai-agentscript
• agent test suites and coverage loops → sf-ai-agentforce-testing
• persona / voice design → sf-ai-agentforce-persona

Required Context to Gather First

Ask for or infer:

• whether this is a Builder / Setup UI project or a code-first Agent Script project
• whether the user is editing Builder metadata or a .agent authoring bundle
• agent type: Service Agent or Employee Agent
• whether the work targets topics, actions, Prompt Builder templates, Models API, or custom Lightning types
• what supporting Flow / Apex / metadata dependencies already exist
• whether the user needs authoring help, publish help, or troubleshooting

Two Agentforce Paths

If the user is starting from scratch and wants strong control over flow/state, route to Agent Script.

Builder Workflow Summary

1. Confirm this is a Builder / Setup UI project
2. Pick Service Agent vs Employee Agent
3. For Service Agents, provision the running user (prefer sf org create agent-user)
4. For Employee Agents, plan visibility with a Permission Set containing <agentAccesses>
5. Define topics with strong descriptions, scope, and instructions
6. Prepare supporting actions (Flow, Apex, Prompt Builder template)
7. Configure inputs / outputs carefully
8. Validate dependencies and template status
9. Publish, then activate

Expanded workflow: references/builder-workflow.md

Key Platform Rules

Topic quality matters

Topic descriptions are routing instructions for the planner. They must be:

• specific
• scenario-based
• non-overlapping with sibling topics

Actions are wrappers around real targets

Prompt Template vs GenAiPromptTemplate

• Prompt Template is the plain-English / UI term used in Prompt Builder.
• GenAiPromptTemplate is the current Metadata API type for source-driven template work.
• Prefer current source format: genAiPromptTemplates/*.genAiPromptTemplate-meta.xml.
• For flexible Prompt Builder templates, plan around the 5-input maximum and consolidate inputs when needed.
• Prompt content should reference inputs with the current merge-field shape, e.g. {!$Input:TargetRecord} or {!$Input:AdditionalContext}.

Supporting metadata deploys first

Before publishing the agent itself, deploy the supporting stack:

1. metadata / fields if needed
2. Apex if needed
3. Flows if needed
4. GenAiPromptTemplate / GenAiFunction / GenAiPlugin
5. then publish the agent

Service Agent running user

For Service Agents, prefer the native GA command: sf org create agent-user --target-org <alias> --json Use the returned username in the running-user configuration.

Employee Agent visibility

For Employee Agents, ensure end users receive a Permission Set containing <agentAccesses>. Without this, the agent can be active but still invisible in Lightning Experience. See ../sf-permissions/references/agent-access-guide.md.

Publish does not activate

After publish, run sf agent activate separately. For automation, prefer sf agent activate --api-name <AgentName> --version <n> --target-org <alias> --json so the rollout is deterministic and machine-readable.

Metadata Guidance

GenAiFunction

Use when registering a single callable action. Validate:

• target exists
• target is active / deployable
• input names match the target contract
• output names match the target contract
• capability text clearly says when the action should be used

GenAiPlugin

Use when grouping related functions into one logical package.

GenAiPromptTemplate

Use for generated content, not deterministic business rules.

Prefer the current metadata shape:

• metadata type: GenAiPromptTemplate
• folder: genAiPromptTemplates/
• file suffix: .genAiPromptTemplate-meta.xml
• content lives under templateVersions
• use published template versions before wiring actions that depend on them

Models API

Use when the solution belongs in Apex-driven AI orchestration rather than Builder-only actions.

Custom Lightning Types

Use when the action needs richer structured input or output presentation.

Expanded references:

• references/metadata-reference.md
• references/genaiprompttemplate.md

Cross-Skill Integration

Recommended Orchestration Order

sf-metadata → sf-apex → sf-flow → sf-ai-agentforce → sf-deploy

Required delegations

High-Signal Failure Patterns

Reference Map

Start here

• references/builder-workflow.md
• references/metadata-reference.md
• references/genaiprompttemplate.md
• references/cli-commands.md

Terminology and template planning

• references/prompt-templates.md
• references/models-api.md
• references/custom-lightning-types.md

Rubric

• references/scoring-rubric.md

Cross-skill reads

• sf-ai-agentscript
• sf-ai-agentforce-testing
• sf-flow
• sf-apex
• sf-permissions
• sf-deploy

Score Guide

Full rubric: references/scoring-rubric.md