Reasoning Budgets & Efficiency Frontiers

Definition

An agent compute budget bounds resources such as model calls, generated tokens, tool attempts, elapsed time and delegated workers. An efficiency frontier compares designs that offer different trade-offs among cost, accepted outcomes and latency.

Efficiency is conditional on the acceptance standard. A cheaper system that fails required safety checks is not an eligible alternative, even if its average task score is attractive. Establish non-negotiable constraints before comparing resource trade-offs.

What it covers

Envelope

Set limits for each task and its delegated children.

Progress

Distinguish new evidence from repeated unsuccessful behavior.

Comparison

Measure quality, cost and time on matched cases.

Stopping

Define success, escalation and incomplete outcomes.

Measure several budgets because they capture different failure modes. A token limit does not bound a costly external API, and a call-count limit does not bound a long-running job. Parent and child limits should share an accounting model so delegation cannot evade the approved envelope.

In practice

Practical design patterns by Min.

• Compare a simple baseline with added reflection, search or parallel attempts.
• Measure the extra accepted outcomes obtained for the extra resources.
• Enforce limits outside the model and reserve room for safe reporting or recovery.

Use controlled comparisons to locate waste. Disable one optional component at a time while keeping task data and evaluation conditions comparable. Record stochastic variation; a small score difference from a single run is weak evidence for a permanent architectural decision.

Research notes

Source-grounded details and editorial applications. Follow each link for the original document.

Working example

Why the controls matter

The illustrative comparison measures marginal benefit, not a universal price for success. If the five recovered cases are consequential, the added expense may be justified. If B also introduces prohibited actions, its higher pass count does not compensate for that failure.

Common pitfalls

• Assuming more reasoning always improves results.
• Letting each subagent receive a fresh unlimited budget.
• Stopping without reporting incomplete work or unresolved side effects.

Termination should leave an understandable state. The agent can return verified partial findings, explain what remains unresolved and request further authorization where appropriate. It should not convert exhausted resources into a fabricated success claim to satisfy the objective.

Questions for operational review

Use these questions to decide what evidence your team needs. They are editorial prompts, not a compliance checklist.

• Which additional steps produce measurable benefit?
• Do child tasks share the parent’s resource envelope?
• How is incomplete work represented at the limit?