The Customer Is Always Right (but not always human) (sergey.substack.com)

OpenAI released GPT-5-codex, a GPT-5 variant tuned for coding that — beyond benchmark scores — demonstrated a striking capability: it ran uninterrupted for seven hours on complex tasks during testing. That extended continuous work is a step-change for autonomous AI agents: instead of short back-and-forth interactions, agents can pursue long-horizon projects, request asynchronous feedback, and chain planning, research, coding and debugging over many GPU hours. Practically, this magnifies compute and cost pressures (modern GPUs cost ~$3–$15/hr and frontier models use multiple GPUs), so subscription tiers alone won’t economically support large, continuous agent workloads. The consequence is an emergent agent economy: token-based “thinking” budgets already map to dollars in APIs, and agents will learn to allocate tokens optimally (e.g., split planning/research/coding/debugging percentages). Beyond internal compute, agents will spend money on external operations — database queries, paid data, specialized generative models, hosting, and web access (Cloudflare is pushing micro-payments for crawlers). Google’s Agent Payments Protocol and stablecoin support accelerate tiny payments between agents. For ML and systems researchers this means new priorities: budget-aware planning algorithms, secure identity/payment primitives, billing/attestation APIs, and compliance/safety tooling. A marketplace of services sold to agents (testing, personas, compute, human feedback, on-demand audits) — and the attendant security, fraud, and ethical risks — is now a realistic near-term frontier.