Yazam.ai is a multi-agent LLM pipeline. It turns an immigration applicant's intake answers into a finished, fact-checked E-2 visa business plan, which a lawyer files with USCIS. The plan has two parts: a narrative and a five-year financial model. Several models sit behind the pipeline, and it has a pricing config. This note covers two values: one we keep out of that file, and one we derive from it.

Why shouldn't thinking-token budgets live in a pricing config?

A price and a budget answer different questions. A price is a fact about a vendor. A thinking budget is a decision about the pipeline: how much reasoning an agent is allowed.

The Yazam.ai code states this distinction directly. A thinking budget is pipeline policy, not pricing data. For that reason, the budgets are defined in code instead of in the pricing config.

The hazard is that the two kinds of value look identical in a config file. An integer next to a model name could be a rate, a limit or a reasoning allowance. The file itself does not say which one it is, or who should be changing it.

What are the three thinking budget tiers?

TierThinking budget (tokens)
max16000
high10000
medium6000

Written compactly, that is 16k/10k/6k. The tier names describe an amount of reasoning. None of them refers to a price.

What could go wrong if policy values and price values shared a file?

The scenario below illustrates the design rationale. It is a hypothetical, not an incident we are reporting from Yazam.ai.

Imagine reasoning allowances stored in the same file as prices, and someone swapping a model. A model swap is a legitimate pricing edit. With the allowances sitting nearby, the person making that edit could also end up changing how much reasoning each agent gets. They might:

In that scenario, the diff could look like a price update and be reviewed as one, even though it changes pipeline behaviour.

The pressure can also run the other way. Someone focused on cost could see a reasoning allowance in a pricing file as a cost lever rather than a quality decision.

Putting the budgets in code is meant to keep these two decisions apart. Editing pricing data is not where reasoning allowances get set, and changing a budget means changing code. That is the argument for the rule: it is designed to pay off the first time a model is swapped.

Why is the agentic financial-modeling loop switched from config instead?

The module that defines the budgets also derives a boolean feature flag from the pricing config. Its purpose is operational. It lets the agentic financial-modeling loop be reverted in production without a code change.

This is consistent with the rule, not an exception to it. The rule is not "config holds data, code holds everything else." The rule is that each value lives where its change process belongs:

How do you decide whether a value is policy or pricing?

These questions are practitioner guidance, not a description of Yazam.ai internals:

Applied here, the thinking budgets answer yes to the second question, so they live in code. The financial-modeling revert switch answers yes to the third, so it is derived from config. In general, neither placement is about convenience. Both come down to who should be able to change a value, and through which path.