The Token Efficiency Playbook — The 5 Harness Rules OpenAI Runs Its Own Agents On
The companion to the video. On July 29, 2026 OpenAI published 'How GPT-5.6 fuses frontier intelligence with frontier efficiency' — and inside it are the 5 harness rules OpenAI runs its OWN agents on. THIS PLAYBOOK: (1) THE ONE DIAGRAM that explains agent costs — one user turn = many model+tool loops, and any cost inside the loop is paid every trip. (2) THE 5 RULES with a 'do this' for each — defer tool discovery (load tools, skills, and plugins only when the task needs them), cap tool output (OpenAI's own default: 10,000 tokens per tool result), append-only context (never insert or edit earlier messages), deterministic tool ordering, and volatile state (timestamps, approval policies, session info) OUT of tool definitions and early context — apply it at execution time. (3) THE PROMPT-CACHING CHEAT SHEET for BOTH providers, verified against the official docs on July 30, 2026 — OpenAI's automatic prefix caching (kicks in at 1,024 tokens, cached input ~10% of the input price — gpt-5.6-luna $1.00 input / $0.10 cached, cache writes 1.25x, 30-minute TTL, prompt_cache_key) next to Anthropic's explicit cache_control breakpoints (up to 4 per request, write premium 1.25x or 2x for the 1-hour TTL, cache reads 0.1x, 512-token minimum on the newest models). (4) OUR REAL RECEIPTS — the same 9.7K-token agent prompt run 4 ways against the live API: append-only context got 6,275 of ~9.7K tokens served from cache (64%) and cut input cost 58%, while the identical prompt with ONE timestamp at the top of the system message got 0 cached tokens. The full 4-row table is inside. (5) THE 10-POINT AGENT AUDIT — the 5 rules turned into checkable boxes you can run on the agent you already have tonight. Independent playbook from Hyperautomation Labs — not affiliated with OpenAI or Anthropic.
Free. No spam. Unsubscribe anytime.