การใช้โทเคนและการตรวจสอบค่าใช้จ่าย

Agents that loop, retry, or use long context can quietly become expensive. Without visibility you only discover the bill at month end.

Token and cost monitoring turns spend into a metric you can watch, alert on, and optimize.

Every call splits into:

• Prompt tokens: everything you send in (system prompt, context, history)
• Completion tokens: what the model generates

They are usually priced differently, so track them separately.

LangChain exposes usage via callbacks. get_openai_callback aggregates tokens and cost for everything inside its context block.

The callback object also exposes the split, which is what you log per request.

Many chat models also attach a usage_metadata field to the response, useful when you call the model directly without a callback.

To stay under a budget, estimate tokens before calling the model using a tokenizer like tiktoken. This catches oversized prompts early.

Define a per-request and per-user token budget. If an estimate exceeds it, trim context, lower k in retrieval, or reject the request before paying for it.

Agents make many sub-calls: tool selection, tool output processing, final answer. Wrapping each step's callback shows which step dominates cost, so you optimize the right one.

Emit tokens and cost as structured logs or metrics (e.g. to Prometheus or LangSmith). Tag them with user, model, and route so you can slice spend.

Once you can see spend, the biggest wins are usually:

• Smaller models for simple steps
• Caching repeated calls
• Trimming history and retrieved context
• Stopping runaway agent loops with iteration limits

Set alerts for sudden cost spikes — often a sign of a prompt-injection loop or a misbehaving tool. Catching it in minutes beats finding it on the invoice.

Test your cost monitoring knowledge.

You learned to observe agent spend:

• Split prompt vs completion tokens
• Use get_openai_callback and usage_metadata
• Estimate with tiktoken and enforce budgets
• Attribute cost per agent step
• Log to dashboards and alert on spikes

Visibility into cost is the foundation for optimizing production agents.