Lesson 13 of 14 · 3 min

Glossary

Every term-of-art used across the tutorial, with one-line definitions.

Agent Script: the DSL inside .agent files. Indentation- sensitive, not YAML.
Bundle / AiAuthoringBundle: the metadata wrapper holding a .agent file plus a bundle-meta.xml.
Topic / subagent: a named scope holding instructions, actions, and state. The planner is "in" exactly one at a time.
Action target: the URI string identifying the backing logic. apex://ClassName, flow://FlowApiName, prompt://TemplateName.
Invocable: Apex annotated with @InvocableMethod and @InvocableVariable. List-in / list-out shape required.
Level 1 vs Level 2 actions: definition (target + schema) vs. invocation (when + bindings).
with X = ...: slot-fill. The LLM extracts X from conversation.
with X = @variables.Y: direct binding. Deterministic.
available when <expr>: visibility gate. The action is hidden from the LLM when the expression is false.
set @variables.X = @outputs.Y: capture an action output to a variable. Captures whatever the planner reports the output to be.
after_reasoning:: post-action deterministic state update block.
@system_variables.user_input: the raw current user message.
Agent running user: the dedicated service user the planner executes as. Needs permset access to every Apex class your actions target.
Trace / plan_id.json: the per-turn JSON record of every step the planner took. The single most useful debugging artifact.
Neuro-symbolic / Type-2 hybrid: Kautz's category for systems where a symbolic problem solver invokes a neural subroutine. Agent Script's closest formal home in the literature.
Constrained generation / guided generation: research line on restricting LLM output to a grammar or FSM. The closest formal analogue to how available when shapes the LLM's tool-choice distribution.