G10 — How the agent loop works

• Publisher: Anthropic / Claude Code Docs
• Primary source: How the agent loop works
• Observed: 2026-10-22 (Europe/Berlin session date)
• Publication date: unknown — no reliable date established from the read
• Updated date: unknown — no reliable date established from the read
• Read status: opened and relevant section read through web tool
• Evidence locator: Control how the loop runs — Turns and budget; Budget headroom
• Evidence class: primary maintainer/provider guidance; not an independent impact study

Short evidence fragment

Spend can therefore pass the cap by up to the cost of that one response

Researcher’s reading note

The SDK exposes max turns and estimated-spend caps. The documentation describes subagent inclusion and post-response checks, so a cap can overshoot.

Limits and use

Version-dependent behavior: documented cap enforcement requires Claude Code v2.1.217 or later. This is one SDK example; external tools and infrastructure need separate control.

Retention

This card preserves a short evidence fragment and an original reading note; full source text is not copied. FinOps/FOCUS adaptations should credit the FinOps Foundation / FOCUS as applicable (their pages state CC BY 4.0). Re-open mutable documentation before implementing controls.