AI Agent Token Usage: Understand Context Growth and LLM Costs

Input tokens vs. output tokens

Every LLM API call involves two distinct token counts:

Agent Analyzer sums both across all steps:

// From analyzer.ts — calculateTokenUsage
for (const s of steps) {
  totalInputTokens += s.inputTokens ?? 0
  totalOutputTokens += s.outputTokens ?? 0
}
totalTokens = totalInputTokens + totalOutputTokens

Why context grows across LLM calls

In a typical ReAct-style agent loop, the model receives the full conversation and tool results on every call. As the agent executes more steps — retrieving data, calling tools, receiving results — each new piece of information is appended to the context that gets sent on the next LLM call.

The result: input tokens grow monotonically. The fifth LLM call in an agent run is often 3–10× more expensive per token count than the first.

Context growth detector

Agent Analyzer compares the inputTokens of the first and last LLM steps in the trace. If the ratio is 3× or more, it flags a context growth issue:

// From analyzer.ts — detectContextGrowth
const first = llmSteps[0].inputTokens
const last  = llmSteps[llmSteps.length - 1].inputTokens
const ratio = last / first

if (ratio >= 3) {
  severity = ratio >= 5 ? 'high' : 'medium'
  // "Input tokens grew X× from first to last LLM call"
}

Token growth visualization

The results page includes a cumulative input token chart that plots inputTokens for each LLM step in order. This makes it easy to see how steeply context is growing and at which step the growth accelerates. Tool steps and retrieval steps are excluded from this chart since they do not contribute input tokens to subsequent LLM calls directly.

Reducing token usage

When context growth is flagged, the issue recommendation suggests:

• →Summarize context between LLM calls rather than appending full results
• →Use a sliding-window strategy that truncates older context
• →Filter tool outputs to include only the relevant fields before passing to the model
• →Split the workflow into smaller sub-agents with limited context scope

For cost implications see AI agent cost analysis. For latency implications see AI agent latency.