Engineers no longer write prompts for AI. They design self-running loops that let agents work autonomously — a shift that has compressed four distinct coding eras into three years.
Back
Loop Engineering replaces prompts as AI coding shifts 4 times in 3 years
Engineers no longer write prompts for AI. They design self-running loops that let agents work autonomously — a shift that has compressed four distinct coding eras into three years.
Nvidia's Jensen Huang declared prompt writing obsolete, marking the arrival of Loop Engineering — a method where engineers design autonomous feedback systems instead of typing instructions, compressing four distinct AI coding eras into three years.
"Nobody writes prompts anymore," Huang said. "The new job is to write and handle loops." Boris Cherny, creator of Anthropic's Claude Code, described a similar shift: "I don't write prompts for Claude anymore. I have loops running that tell Claude what to do and decide what happens next."
Claude Code now ships three loop primitives — /loop for timed cycles, /goal for objective-driven execution until verification passes, and /schedule for cloud-based unattended runs. The /goal command enforces a critical architectural rule: the model writing code cannot verify its own output. Anthropic routes code generation through its large model while a separate, smaller Haiku model handles acceptance testing. OpenAI's Codex takes a parallel approach, spinning up as many as eight agents in isolated cloud sandboxes, each working on a subtask before merging results.
The shift from prompts to loops carries direct implications for AI infrastructure spending. Autonomous loop systems consume more tokens per task than single-prompt workflows, increasing demand for inference compute. Nvidia stands to benefit as the primary supplier of GPUs powering these multi-agent loops, while cloud providers — Amazon Web Services, Microsoft Azure, and Google Cloud — compete for the incremental workload. Anthropic, which filed for an IPO last week, has embedded loop architecture directly into Claude Code, potentially accelerating enterprise adoption and recurring API revenue.
The four-stage evolution traces a clear trajectory. From 2023 to 2024, Prompt Engineering dominated — users crafted precise instructions for each interaction, and output quality depended entirely on prompt craftsmanship. Around 2024 to 2025, Context Engineering shifted the focus from "how to ask" to "what to show the model," as RAG pipelines and codebase integration expanded the information available to each query. By 2025 to 2026, Harness Engineering emerged, giving agents access to tools, APIs, and real-world execution environments. Loop Engineering, the fourth and current phase, closes the cycle by enabling agents to run autonomously — planning, executing, verifying, and retrying without human intervention at each step.
The Architecture Behind the Loop
The core insight is separation of concerns. In Claude Code's /goal system, one model generates code while an independent verifier — a separate model with no visibility into the generator's reasoning — tests the output. This prevents the "grading your own homework" problem that plagues single-model workflows. The verifier rejects outputs that look plausible but fail functional tests, forcing the generator back into the loop.
Google engineering lead Addy Osmani, who coined the term Loop Engineering, struck a cautious tone. "Still early. I'm reserving judgment. You have to be very careful about token costs," he wrote. The warning is not theoretical: loop systems that lack hard stop conditions — token limits, iteration caps, or time boundaries — can run until they exhaust budgets or hit API rate limits.
The 'Understanding Debt' Problem
Andrej Karpathy, speaking at Sequoia Capital's AI Ascent 2026 conference, offered a counterpoint to the automation enthusiasm. "You can outsource your thinking, but you can't outsource your understanding," he said, citing a line he said he has revisited repeatedly. As loops merge code faster than humans can review it, engineers accumulate "understanding debt" — systems that no one on the team fully comprehends. The practical cost is not the token bill, Karpathy suggested, but the day someone must debug a system nobody has read.
Cherny, who said he deleted his IDE in November 2025 and now manages hundreds of agents from his phone, acknowledged the trade-off. Agents that cannot resolve issues escalate to his inbox. His workflow represents the endpoint of the loop method: humans write rules and make judgment calls; agents execute everything else.
The academic foundation for loop systems traces to Yao Shunyu's 2022 ReAct framework (Reason + Act), which received an ICLR 2023 Oral designation and has accumulated tens of thousands of citations. ReAct formalized the think-act-observe cycle that underlies every modern agent loop. Subsequent work — Reflexion's error-feedback mechanism, Tree of Thoughts' multi-path search, and a cascade of tool-use agent papers — converged into the engineering discipline now called Loop Engineering.
For investors, the key metric is cost per accepted change. If a loop's acceptance rate falls below 50%, the system is losing money — the human is doing the review work the loop was supposed to automate. As enterprises deploy loop architectures at scale, the winners will be those that minimize token waste while maximizing autonomous throughput.
This article is for informational purposes only and does not constitute investment advice.
