Skip to main content
Work from totals to mechanisms. For each finding, record evidence, size, confidence, and an alternative explanation.

​
1. Validate attribution

Check:
  • provider status and last-updated timestamp;
  • actual, allocated, measured, and unallocated totals;
  • token/request coverage;
  • currency and billing period;
  • whether the target project is isolated from shared provider usage;
  • model/Agent/task dimension coverage and collector truncation warnings;
  • whether scheduler statuses represent enqueue records or final task outcomes.
A large unallocated share or stale or missing provider data prevents a confident project-level conclusion. Report this limitation before discussing optimization. If model coverage is low, report Unknown model as a measurement gap. Do not assign historical cost using an Agent’s current model configuration.

​
2. Find the change point

Compare daily cost, requests, and tokens. Align spikes with:
  • model changes;
  • schedule creation or interval changes;
  • prompt/skill deployment;
  • new data sources or file formats;
  • retry/error periods;
  • new Agents, Hooks, or follow-ups.
Use equal windows around the change. Normalize by successful business outcome when possible.

​
3. Separate model price from volume

For each model calculate cost share, request share, token share, and effective cost per request. Then compare current official rates. Evidence for a model-price driver:
  • cost share greatly exceeds request/token share;
  • the high-cost model’s active dates align with the spend spike;
  • similarly sized tasks on a lower-cost model cost materially less.
Alternative explanations include longer contexts, more output/reasoning, tool calls, or task complexity concentrated on that model.

​
4. Quantify schedule-required and avoidable volume

Inventory every active schedule and triggering path. Derive intended executions from interval/trigger and compare with observed runs/tasks. Split volume into:
  • required successful executions;
  • retries after errors;
  • duplicate or overlapping triggers;
  • concurrent runs that repeat unfinished work;
  • abandoned or no-op runs;
  • follow-up fan-out.
If schedule volume is a requirement, report its unavoidable baseline cost. Optimize cost per execution and avoidable overhead instead of recommending removal.

​
5. Diagnose prompt and context cost

Use Agent and schedule prompt lengths as a screen, not proof. Use task message summaries to inspect:
  • input growth over time;
  • repeated identical human prompts;
  • largest messages;
  • tool-result size;
  • repeated system/context injection;
  • low cache ratio despite a reusable stable prefix.
Common mechanisms:
  • overly detailed instructions repeated on every request;
  • multiple skills restating the same rules;
  • full conversation replay when a state summary would suffice;
  • entire file sets loaded before the task is narrowed;
  • variable values placed early and invalidating prompt caching.

​
6. Diagnose tool and data amplification

Rank tool-call names in anomalous tasks. Investigate repeated Data Lake/Blob reads, PDF/image extraction, OCR, conversions, searches, and broad directory scans. Ask:
  • Is the same immutable input processed in many runs?
  • Is the derived text/index stored and reused?
  • Can metadata narrow candidate files before reading content?
  • Are large tool results returned to the model when only IDs/counts are needed?
  • Does an error cause the Agent to retry the same expensive operation?
Treat preprocessing and indexing as recommendations only; do not mutate data in a cost-analysis task.

​
7. Diagnose output, reasoning, and loops

High output or reasoning may be appropriate. Look for:
  • repeated long reports when a delta would work;
  • Agent handoffs that duplicate context;
  • evaluator/reviewer loops without a termination condition;
  • tool-call/result cycles with no state change;
  • errors that regenerate an entire answer;
  • model settings mismatched to task difficulty.
Use cost per successful outcome rather than cost per request when quality matters.

​
8. Check budget and operational controls

Report existing project budget policy, thresholds, current usage, and enforcement state. Absence of a budget is a risk-control gap, not a root cause of past spend. Do not change budgets or configuration during analysis.

​
Finding confidence

  • High: provider actuals reconcile and the driver is visible in multiple independent dimensions.
  • Medium: allocation and runtime behavior align, but pricing or task detail is incomplete.
  • Low: based mainly on configuration, token proxy, or a small sample.

​
Recommendation priority

Prioritize by:
Separate:
  • immediate containment;
  • low-risk efficiency;
  • architectural context/data improvements;
  • measurement gaps.
For each recommendation specify the mechanism, affected share, validation experiment, quality guardrail, and rollback condition.