Cost And Token Budget Optimizer For Agent Loops

ROLE: You are a performance engineer optimizing the cost and latency of LLM agent loops.

CONTEXT: My agent does [WORKFLOW]. Current behavior: average [N_STEPS] steps, [TOKENS] tokens per run, model [MODEL]. The biggest cost driver appears to be [SUSPECTED_DRIVER]. Quality must not drop below [QUALITY_BAR].

TASK: Produce an optimization plan.
1. Map where tokens and steps are spent across the loop (context bloat, redundant tool calls, over-long reasoning, re-reading state).
2. Propose targeted optimizations: context pruning/summarization, caching, batching tool calls, cheaper model routing for sub-tasks, and earlier stop conditions.
3. For each optimization, estimate the expected savings and the quality risk.
4. Recommend which sub-tasks can be downgraded to a smaller/cheaper model and which must stay on the strong model.
5. Define a guardrail metric to detect if an optimization silently hurt quality.

OUTPUT FORMAT: A findings list, an optimization table (Change | Est. Savings | Quality Risk | Effort), a recommended rollout order, and the quality guardrail metric.

CONSTRAINTS: Never trade away correctness to save tokens below [QUALITY_BAR]. Prefer reversible, measurable changes. Be specific about where in the loop each change applies.