Skip to main content

Context Optimization With use one of

By now, every tutorial has used use to decide what the model can see. use one of makes that decision explicit as a set of named alternatives.

Context Optimization with use one of tutorial overview

This is useful when you know a context slot matters, but you want to compare different policies:

  • concise vs. detailed
  • evidence vs. no evidence
  • cheap context vs. rich context
  • default behavior vs. escalation behavior

Full source: ../../../tutorials/use-one-of.as

1. The Complete Program​

Create use-one-of.as, or open the repository copy at tutorials/use-one-of.as:

import llm Qwen from "ollama://localhost:11434/qwen3.6"

main agent ContextChoice {
    model Qwen
    role "Support triage assistant"
    description "Show how one context slot can expose alternative context policies."

    main func(input {
        request: string
        customer_tier: string
    }) {
        policy = {
            concise: "Use a short answer for routine requests.",
            detailed: "Explain reasoning and give concrete next steps.",
            vip: "Prioritize escalation paths and account impact."
        }

        use input.request as "support request"

        use one of {
            concise: policy.concise selected
            detailed: policy.detailed
            vip: policy.vip
            none: empty
        } as "response policy"

        use input.customer_tier as "customer tier"

        generate({ input: "Triage the support request", max_output: 500 }) -> {
            priority
            summary
            next_steps: list[string]
        }
    }
}

2. One Slot, Multiple Candidates​

This block declares one context slot named "response policy":

use one of {
    concise: policy.concise selected
    detailed: policy.detailed
    vip: policy.vip
    none: empty
} as "response policy"

Only one candidate is active at a time. In the source above, concise is marked selected, so it is the default.

3. Why Name the Candidates?​

Variant names are a contract between you, traces, and optimizer tooling. A trace can say "the concise variant was picked" instead of only showing an anonymous chunk of prompt text.

Good names describe the policy:

  • concise
  • detailed
  • vip
  • none

The none: empty candidate means this context slot can be omitted entirely.

4. Run It​

Run with mock output:

agentscript tutorials/use-one-of.as --mock --input '{"request":"Checkout is failing for a customer","customer_tier":"vip"}'

Print the trace to see which variant was picked:

agentscript tutorials/use-one-of.as --mock --trace --input '{"request":"Checkout is failing for a customer","customer_tier":"vip"}'

5. Where Optimization Fits​

use one of does not automatically search by itself. It exposes a stable search space: named context choices with stable site ids.

For an end-to-end optimizer example, see:

That example uses host://optimizer to inspect variant sites, run trials, and preview a specialization.