Token Strategy — Efficiency, Balance, Aggressive, and Staffing

How will you spend tokens?

If business asks "does the loop earn," the token strategy asks "how do you spend tokens." Saving tokens and burning them are both strategies — which is right is decided by the business situation of the moment.

You can run the same loop in three modes. Efficiency minimizes tokens, balance differentiates by path, and aggressive spares no tokens. A mode is not a permanent choice but an operating lever you switch as the situation changes.

Efficiency — the saving strategy

Minimize token consumption. Press unit cost to the floor with lighter models, aggressive caching, short context, and batching.

Choose efficiency when margin pressure is high, when handling large volumes of low-risk repetitive traffic, or when turning self-sustaining is urgent. The trade-off is a low ceiling on quality and autonomy — it falls short on paths that need hard judgment or differentiation.

Balance — the differentiating strategy

Differentiate by path. Assign high-performance models to high-risk, high-value decisions and lighter models to low-risk paths like classification and filtering. It is the default for most production.

Choose balance when operating steadily and managing quality and cost at once. The trade-off is that per-path policy and monitoring add operational complexity. Set the boundaries wrong and you miss on both quality and cost.

Aggressive — the burning strategy

Spare no tokens. Deploy top-performance models, extended thinking, and multi-agent verification to push quality and speed up.

Choose aggressive when differentiation is the point, or when high-value decisions and early trust matter more than cost — it fits the phase of buying time with investment. The trade-off is that high token unit cost pressures self-sufficiency. Sustained without margin discipline, cost overtakes revenue.

The mode is called by the business

The three modes are not a ranking but a function of situation. Heading toward self-sufficiency early, press unit cost down with efficiency; in steady operation, catch quality and cost together with balance; in a phase of chasing differentiation with investment, burn tokens aggressively on high-value paths. Even within one product, different modes apply per path — aggressive on the core decision path, efficiency on the peripheral processing path.

Staffing by stage

Choosing a mode alone does not turn the loop. Who runs that mode also changes with the stage. Organize people as roles, not org charts — name at each OCLS stage who owns (OWN), what gets approved (human approval), and where it is tuned (SHARPEN). A role is a point of responsibility on the loop, not a job title, and it requires no reorganization.

At stage one, one person owns direction, execution, and review together. At stage two, designate an outcome owner per role and separate review. At stage three, people running approval gates and evaluation stay resident on the loop. At stage four, validation, operations, and cost management are internalized as roles — still not a reorg but a differentiation of responsibility on the loop.

The validation that holds the strategy up

Quality is not built anew. Evaluation and guardrails (pass@k/pass^k, capability and regression evals), private evals (business-outcome criteria), decision traceability, and governance lint already exist. The token strategy's job is to wire this validation permanently into the loop — whichever mode you pick — so honesty holds whether you burn more tokens or fewer.

Loading diagram…

Saving tokens and burning them are both strategies. Which mode to pick and when is called by the business; who runs that mode is set by the stage. The token strategy is the operating language that ties the two together.

See also

• This theme among the environment's three faces: Token Strategy
• Sister pieces: The Data Flywheel · The Business Economics of the Loop