Model-Agnostic by Construction

Choosing workers and controlling cost

Treat model choice as a task-level decision. Start with the least expensive method that can pass the acceptance test. Use deterministic tools when the transformation is deterministic, bounded small-model work for bulk steps, and deeper reasoning for ambiguous or high-consequence judgments.

A computer-use tool may be required at any reasoning level. Check where it runs and what account it can reach. Changing model or runtime does not preserve access automatically, and it must not expand the audience or authority.

The parent budget covers the whole result: coordination, workers, retries, tools, APIs, review, export and destination readback. Track in-flight commitments as well as completed usage. Reserve enough for verification. Where metering is incomplete, impose bounded steps or time before spending can reach the cap.

Record the current price basis if cost matters. Do not repeat historical savings percentages as a promise. Compare accepted outcomes and rework, not token price alone. Escalate a named uncertainty or failed acceptance test, rather than using a more expensive model by habit.

Route by capability and available capacity

Start with a deterministic tool for a mechanical step. Use a bounded local or lower-cost worker when its quality is sufficient, and deeper reasoning for an ambiguous decision or expensive mistake. Check the actual account's remaining capacity, tools and source access. Historical readings are dated evidence; expired readings are unknown capacity today.

If the preferred account is capped, retain the task, scope and acceptance test while an authorized alternate accepts the next step. This does not authorize overages, subscriptions or a larger budget. Compare total cost and time per accepted result, including review and rework. Local tokens per second measures generation; it does not promise complete audits or autonomous production.