https://andreasnissen.dev/tags/token-efficiency/Recent content in Token Efficiency on Andreas NissenHugoen-us© Andreas NissenMon, 31 Aug 2026 00:00:00 +0000https://andreasnissen.dev/writing/hidden-token-tax-agent-tools/Mon, 31 Aug 2026 00:00:00 +0000https://andreasnissen.dev/writing/hidden-token-tax-agent-tools/<p>The token cost of a tool-using agent starts before it calls a tool.</p> <p>The model first needs to learn which tools exist, what each one does, which inputs are valid, and how their results should be interpreted. It then spends tokens selecting a tool, constructing arguments, reading the result, and deciding what happens next.</p> <p>Optimizing only the final response misses most of this path.</p> <h2 id="the-cost-has-several-layers">The cost has several layers</h2> <p>An agent task can consume model context in at least six places:</p>