Cost Model For Agent Workers

Purpose

Cost control is part of AI-native engineering. Prox OS should not send every task to the most expensive model, and it should not use one giant prompt when a task graph would make the work smaller and safer.

Model Routing

Task Type	Suggested Model Strategy
Idea capture	Cheap / fast model
Prompt compression	Medium model
Architect planning	Strong reasoning model
Scoped implementation	Strong coding agent, small scope
Docs cleanup	Cheaper model
QA output interpretation	Medium model + CI logs
Integration review	Strongest available reasoning/coding model
Business/growth draft	Medium or strong depending on stakes

Principles

Subscription tools are useful for exploration and interactive refinement.
API usage is useful for controlled automation and cost visibility.
Small tasks are cheaper than large tasks because they need less context and retry less often.
Hard scope lowers token cost, failure cost, review cost, and merge risk.
Important architecture and final integration reviews deserve stronger models.
Low-risk cleanup and first-pass summarization can use cheaper models.
Human review remains the final cost-control and risk-control layer.

Cost Gates

For opt-in control-plane or multi-agent work:

Classify task size with docs/development/ai/protocols/cost-gate.md.
Decide if Architect AI, RFC, ADR, or task graph is needed.
Set allowed paths and forbidden paths.
Pick the cheapest model class that can safely complete the task.
Define validation before implementation begins.

Avoid

Do not optimize cost by bypassing platform rules or access controls.
Do not hide work in one huge "do everything" prompt.
Do not let cheap first-pass output become accepted architecture without review.
Do not run multiple agents on the same files without an integration plan.