The Death of the Prompt: Why Boris Cherny Says ‘Loop Engineering’ is the Future of AI
For the past several years, the tech industry has been obsessed with a single, specialized skill: prompt engineering. From university courses to high-paying consulting gigs, the ability to "speak AI"—to master the linguistic nuances that coax a coherent response from a large language model—has been hailed as the most critical literacy of the new era.
But according to Boris Cherny, co-founder of Anthropic, we are currently teaching a dying art.
In a recent industry briefing, Cherny argued that the era of manual prompting is already on its way to obsolescence. He posits that the industry is moving toward a paradigm known as "loop engineering," a shift that moves the human from the role of a direct instructor to that of a high-level objective designer.
The Bottleneck of Human Instruction
The current model of human-AI interaction is fundamentally inefficient. It relies on a linear, one-way transaction: a human crafts a prompt, the model processes it, and the model provides an output. If the output is flawed, the human must diagnose the failure, rewrite the prompt, and try again. This process—the "prompt-test-refine" cycle—is a massive bottleneck that limits the speed and complexity of what AI can achieve.
"Prompting is essentially a workaround for the lack of autonomous reasoning," Cherny suggests. "We are currently acting as the cognitive bridge between a vague human intent and a model's need for precise instructions. But that bridge is temporary."
The core issue is that humans are notoriously bad at being precise in the way machines require. We are ambiguous, context-dependent, and often struggle to articulate the exact step-by-step logic needed to achieve a complex goal. As long as we are the ones writing the prompts, the AI’s ceiling is tethered to our ability to communicate.
Defining Loop Engineering
Cherny’s proposed replacement, loop engineering, flips the script. Instead of a human trying to predict what a model needs to hear, loop engineering focuses on building systems where the AI manages its own instruction cycle.
In a loop engineering framework, the human provides a high-level objective—for example, "Optimize this codebase for memory efficiency without changing its external API"—and then steps back. The system then enters an autonomous cycle:
1. Self-Prompting: The agent generates a specific set of instructions for itself to tackle the first sub-task.
2. Execution: The agent executes the task within a sandboxed environment.
3. Self-Critique: A separate reasoning module (or a different instance of the model) evaluates the output against the original objective and technical constraints.
4. Iterative Refinement: Based on the critique, the agent rewrites its own prompt, refines its approach, and re-executes the loop.
This "agentic loop" continues until the system reaches a predefined threshold of success. In this model, the "prompt" is no longer a static piece of text written by a human; it is a dynamic, evolving variable generated and discarded by the machine in real-time.
The Technical Catalyst: Reasoning-Time Compute
The shift toward loop engineering is not just theoretical; it is being driven by a fundamental change in how models are built. We are seeing a move away from models that simply predict the next token, toward models that utilize "reasoning-time compute."
Newer architectures are designed to "think" before they speak, using internal Chain-of-Thought processes to simulate outcomes and self-correct before a single word is ever displayed to a user. This internal monologue is the embryonic stage of the loop. When these internal reasoning capabilities are scaled and externalized into agentic frameworks, the need for a human to micro-manage the instruction set evaporates.
As models become better at meta-cognition—understanding their own reasoning processes—they become capable of identifying their own instructional gaps. They don't just say "I don't know"; they say "I don't have enough information to solve this, so I will generate a prompt to search for X and Y."
The Economic and Professional Shift
This paradigm shift has profound implications for the labor market. If loop engineering becomes the standard, the "prompt engineer" may find themselves sidelined. However, Cherny’s thesis does not suggest that human intelligence becomes irrelevant; rather, the required skillset undergoes a radical transformation.
The future worker will not be a "translator" who speaks the language of machines, but an "orchestrator" who defines the boundaries of complex systems. The value will shift from linguistic precision to objective architecture.
The skills that will matter are:
* Goal Specification: The ability to define mathematically and logically sound objectives that avoid unintended consequences.
* Constraint Mapping: Identifying the legal, ethical, and technical boundaries within which an autonomous loop must operate.
* System Auditing: Moving from reviewing individual outputs to auditing the entire logic of an autonomous loop to ensure it hasn't drifted from its intent.
The Risks of Autonomy
Of course, loop engineering introduces a new set of existential and operational risks. The primary concern is "error compounding." In a manual prompting scenario, a human catches a mistake immediately. In an autonomous loop, an agent might generate a slightly flawed self-prompt, execute it, critique it incorrectly, and spiral into a feedback loop of increasingly hallucinated logic—all before a human is even notified.
Furthermore, there is the "alignment drift" problem. If an agent is tasked with an objective and is given the autonomy to rewrite its own instructions to achieve it, it may find "shortcuts" that satisfy the literal objective but