The discipline of reducing the tokens consumed per user-visible outcome in an agentic AI system, without sacrificing capability.
AI agent token optimization is the practice of analysing where tokens are spent across the lifecycle of a single agentic task — including model narration, repeated context, retries from format failures, free-form generation of structured data, and the system prompt itself — and reducing each category through targeted architectural and prompting changes.
In a typical unoptimised agent system, only about 8% of total tokens are the actual user-visible answer. The other 92% is the cost of how the system thinks, not what it thinks — repeated context, self-narration, retry storms, verbose system prompts, and free-form generation of data the model could have emitted as a typed schema. This is where order-of-magnitude reductions live.
Across Skopa engagements, agentic systems typically arrive at 10:1 to 30:1 ratio of LLM calls to user-visible outcomes. After a disciplined token optimisation engagement they settle at 2:1 to 5:1 with no quality loss. The levers that matter: separating planning from execution, pushing deterministic logic into code, replacing free-form output with typed schemas, removing self-narration, compressing the standing context, caching the stable prefix, bounded iteration with hard caps. Each is unremarkable in isolation. Compounded, they routinely produce reductions in the 8×–20× range.