agent cost governor

The problem

LLM bills scale faster than utility, but manual instrumentation for cost control is error-prone. Developers often waste money on redundant prompt tokens or risk application stability with poorly implemented budget halts.

What it does

• Generates prompt caching cache_control wiring at static content breakpoints.
• Implements model-tier downgrade proposals for low-complexity tasks like classification or extraction.
• Produces budget guards for agent loops with flag-and-continue alerting and backstop ceilings.
• Instruments cost-per-call logging that captures token usage without leaking prompt content.

Why this beats prompting it yourself

This skill prevents "fixer" hallucinations like placing cache breakpoints below a model's minimum token threshold, which silently fails. It enforces a strict safety protocol for budget guards, ensuring you don't trade production availability for cost savings by accident.

Use cases

• Add Anthropic prompt caching to a high-volume RAG pipeline.
• Downgrade specific extraction tasks from Sonnet to Haiku based on task complexity.
• Wire a budget governor into a long-running autonomous agent loop.
• Implement granular cost-per-call logging across a multi-model stack.

Known limitations

Does not perform initial cost audits or ROI estimation. Not for managing financial transaction risk or credential safety in money-moving agents.