sf-ai-agentscript_skill

Overview

This skill enables development of deterministic Agentforce agents using the Agent Script DSL in a single .agent file. It covers syntax, debugging, testing, and CLI deployment for FSM-style agents with instruction resolution and hybrid reasoning. The skill highlights compile-time constraints, preview quirks, and production pitfalls so agents behave predictably in orgs running API v65.0+.

How this skill works

The skill inspects Agent Script source for required structure (one start_agent, folder/agent_name match), DSL syntax rules, and reserved names that cause parse errors. It validates lifecycle hooks (before_reasoning/after_reasoning), action definitions (complex_data_type_name mapping), and post-action directives allowed only on @actions.*. It also documents Preview/Canvas corruption bugs and provides patterns to test linked variables and outputs safely.

When to use it

• Authoring new Agentforce agents with deterministic behavior
• Debugging compile errors caused by reserved names or syntax misuse
• Preparing agents for production to meet API, license, and user prerequisites
• Testing agent behavior in both Preview and Live modes before deployment
• Optimizing credit consumption by moving data fetch to lifecycle hooks

Best practices

• Keep all structural edits in Script view; use Canvas only for light visual checks
• Enforce consistent indentation (no mixed tabs/spaces) and capitalize booleans (True/False)
• Use compound conditions or flattened sequential ifs—never use else if or nested if blocks
• Place only one start_agent block and ensure agent_name exactly matches folder name (case-sensitive)
• Use before_reasoning to fetch/cache data and after_reasoning for logging to reduce runtime credits
• Define action complex_data_type_name via the UI to avoid manual mapping errors

Example use cases

• Create an FSM agent that uses before_reasoning to fetch account data once, then reuses variables across topics
• Build a demo helper topic to set state without backend actions for local testing
• Convert prompt-heavy behavior into code-enforced transitions and deterministic post-action logic
• Validate and fix Canvas-related corruption by toggling edits in Script view and rechecking in Preview
• Deploy agent via CLI after verifying default_agent_user exists and required API/license prerequisites