On August 5, 2026, Jeff Dean announced he was leaving Google — and with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, founding a company called Discovery Loop. The press coverage fixated on the departure of Google’s AI core. But the signal worth reading was buried in the name itself: Loop.
That was no accidental branding. Over the past few months, “the Loop” has evolved from workflow slang among a small group of developers into the most consequential concept in AI engineering. Google engineer Addy Osmani systematized the approach in an essay titled Loop Engineering. IBM formally defined it as the engineering discipline of designing agent workflows that act, observe, decide, and iterate. Boris Cherny, who leads Claude Code, has said he rarely prompts Claude directly anymore — “my job is writing Loops.”
From Prompt Engineering to Loop Engineering
Between 2024 and 2025, prompt engineering was the required curriculum for anyone working with AI: how to write system prompts, design few-shot examples, chain reasoning steps. Those techniques formed the basic grammar of human-model interaction.
By 2026 the ground had shifted. Models no longer need a complete instruction to start working. Agent-class products — Claude Code, Cursor, GitHub Copilot — parse tasks on their own, plan steps, execute operations, and verify results. Human input has changed shape: from step-by-step commands to goals and constraints. This is not an incremental improvement; it is a restructure of the division of labor. The relationship between human and AI has moved from operator-and-tool to designer-and-autonomous-system. You no longer tell the model how to do each step. You design a mechanism — a Loop — that lets the model work continuously inside a closed cycle, while you make judgment calls and corrections at the moments that matter.
That is the starting point of Loop Engineering: stop writing prompts; start designing loops.
Anatomy of a Working Loop
Osmani’s essay lays out the elements a production-grade Loop must specify:
- Task trigger: what starts the cycle — a schedule, an event, or a human?
- State persistence: where do intermediate results and context live, and how does the loop resume after interruption?
- Result validation: who judges the output — automated tests or human review?
- Loop conditions: what continues the cycle, and what terminates it?
- Feedback injection: how does each round’s output feed the next round’s input?
Claude Code’s workflow is a textbook Loop: the user states a task, the agent reads the codebase and plans changes, executes them, runs tests, adjusts based on failures, and continues to the next round. The user doesn’t participate in each step — only at checkpoints, like approving the change plan.
Peter Steinberger distilled the discipline into one line: stop prompting coding agents one instruction at a time; design Loops that prompt the agent.
Three Loops, Three Maturity Curves
Closed-loop structures of this kind have now crystallized in three distinct AI domains.
Loop One: Model Research
In March 2026, Andrej Karpathy open-sourced AutoResearch — a system where AI agents modify model-training code, run experiments, check results, and keep the improvements that hold up. The entire environment compresses into one editable training script, one fixed metric, and one GPU, so iterations cycle fast. Karpathy then joined Anthropic to build the team that uses Claude to accelerate pretraining research — a sign that AI-assisted AI research has moved from personal experiments into frontier labs’ core pipelines. Google DeepMind’s AlphaEvolve runs a similar loop: Gemini proposes code and algorithmic approaches, automated evaluators run and score them, and better candidates feed the next round — already applied to datacenter efficiency, chip design, training strategies, and matrix multiplication algorithms.
Loop Two: Scientific Knowledge
Sakana AI’s The AI Scientist, launched in 2024, attempted the full cycle from hypothesis generation through code, experiments, and paper writing; by 2025, an AI-generated paper had cleared peer review at a machine-learning workshop. FutureHouse builds multi-agent systems that connect literature retrieval, data analysis, hypothesis generation, and experiment planning. Discovery Loop aims at the far end of this direction — AI participating in the complete research process: proposing hypotheses, running experiments, evaluating results, and choosing the next line of inquiry from the feedback.
Loop Three: Physical Experiments
Seattle startup Potato is pushing “fully closed-loop autonomous science” from computation into lab robotics; Europe’s Sigmatic Sciences proposes “lab-in-the-loop,” connecting computational models, automated experiments, and data feedback. China has its own entrant — Shanghai’s Academy for AI Science runs Golab, a materials-science R&D factory built on an AI-compute → automated-experiment → data-feedback → model-self-evolution cycle.
The three loops mature at different speeds. Model-research loops run fastest because every modification scores quickly and success or failure is unambiguous. Science loops are starting to produce verifiable results, but feedback depends on expert judgment. Physical loops are slowest — a single validation can involve equipment, samples, and weeks of time.
The Deeper Path: Recursive Self-Improvement
Ordinary Loops let AI keep completing external tasks. A deeper evolution runs underneath — recursive self-improvement, where AI participates in improving its own algorithms, training pipelines, evaluators, or research methods, and applies each round’s improved capability to the next. Not merely “AI writing code” — “AI improving the AI that writes code.”
In May, Yao Tian — formerly director of research science at Meta FAIR — co-founded Recursive Superintelligence with Richard Socher, Tim Rocktäschel, and Jeff Clune, putting the concept in the company name with $650 million raised at a $4.65 billion valuation. In June the company unveiled its first-stage automated AI research system: AI proposes ideas, implements them, runs experiments, validates results, and picks the next research path based on prior experiments. Multiple research branches run in parallel, promising directions merge, and the system checks for reward hacking and experiment variance. The team reports leading results across three benchmarks: fixed-compute model training, small-model training speed, and GPU kernel optimization.
A July 2026 survey that analyzed 1,250 papers on the subject splits current practice into “bounded self-improvement” — where goals, experiment boundaries, and metrics are set by humans, and which has entered industrial use — and “open-ended recursive self-improvement,” which remains constrained by external validation costs, model-collapse risk, and compute limits.
Jeff Dean’s metric from a YC interview captures how the whole field should be measured: “discoveries per unit of compute input.” Not token throughput. Not parameter counts. The conversion rate from compute invested to valid discoveries.
Applying Loop Engineering on Your Own Work
This is not a discipline reserved for large AI labs. Anyone building with AI can start today:
- Move from one-shot tasks to loopable tasks. Audit what you do with AI daily. Any recurring task — code review, bug fixes, test generation — deserves a reusable workflow the AI can cycle through autonomously while you approve at checkpoints.
- Define termination conditions first. The classic Loop failure is running forever. Before designing one, answer three questions: what counts as done? When should a human step in? When should you abandon this approach entirely?
- Build fast, quantifiable evaluation signals. A Loop is only as good as its feedback signal. In code, the test pass rate works. In content generation, you may need to design an automatic scoring mechanism. The vaguer the signal, the weaker the loop — which is exactly why physical-experiment loops progress slowest.
- Level up from writing prompts to writing loop code. As Cherny puts it, his job now is “writing Loops.” The skill set shifts from prompt technique to engineering: defining agent workflows in code, managing state and context, designing validation and rollback, monitoring long-running agents.
- Watch for the value moving up. The deepest change Loop Engineering brings is that human value migrates upward. When you stop issuing per-step instructions, your time reallocates to defining the right goals, designing effective feedback, judging which directions are worth exploring, and correcting course at the key moments.
In his final interview before leaving Google, Jeff Dean offered what may be the single most durable judgment of the Loop Engineering era: “The truly scarce skill of the future is knowing what problems to set agents on.”
The Takeaway
From prompt engineering to loop engineering is not a terminology swap — it is a structural reorganization of the human-machine relationship. In the prompt era, human value meant translating requirements into instructions a model could understand. In the loop era, human value means designing systems, setting goals, defining boundaries, and judging direction.
Jeff Dean spent 27 years helping Google turn data, compute, and model training into scalable systems. Now he and the people who mattered most at Google want to make the scientific method itself into a system that runs at scale. The Loop is their method. Scientific discovery is the first world it enters. Recursive self-improvement may be the road beyond.
For every AI developer, the question is no longer “should I learn loop engineering” — that question has already expired. The real one is: where is your Loop?
Sources
- Tencent Technology, “The most important person at Google left — how important is the Loop he’s building?”
- InfoQ, “Jeff Dean’s final conversation before leaving”
- Addy Osmani, “Loop Engineering”
- IBM definition of act-observe-decide-iterate agent workflows
- Recursive Superintelligence official announcement
- arXiv survey of 1,250 RSI papers
