Lesson 3 of 14 · 5 min

The mental model: four control surfaces

Topic boundaries, instructions, available-when gates, and after_reasoning. Knowing which to reach for first is the whole skill.

Agent Script gives you four orthogonal places to enforce behavior. Knowing which to reach for first is the whole skill.

Surface	What it controls	When to use
Topic / subagent boundaries	Which set of actions and instructions is "active"	Coarse phase changes (verification then refund)
`reasoning.instructions:`	What the LLM is told to do this turn	Soft guidance, multi-step plans, polite phrasing
`available when` gates	Which actions the LLM can even see	Hard constraints (no refund without verification)
`after_reasoning` blocks	Variable updates after an action runs	Compute derived state from action outputs

Worked example

subagent identity_verification:
    actions:
        verify_customer:
            target: "apex://VerifyCustomer"   # backing action (truth source)

    reasoning:
        instructions: ->                       # prompt (LLM-driven)
            if @variables.customer_verified == False:
                | Ask for an email if you don't have one. Otherwise call verify.
        actions:
            verify: @actions.verify_customer
            go_refund: @utils.transition to @subagent.refund_processing
                available when @variables.customer_verified == True   # gate

    after_reasoning:                           # derived state
        if @variables.verified_flag == "true":
            set @variables.customer_verified = True

Each line uses a different surface. The LLM is in charge inside instructions: but cannot bypass available when or after_reasoning.

Quick check

The agent must never issue a refund without first verifying the customer. Where do you enforce this?