The Deep Feed

01 — Lenny's Newsletter

The Loop Engineering Revolution

Moving beyond simple prompts to autonomous goal-seeking agents

By Lenny Rachitsky · 12 min read

Editor's note: Understanding the difference between a script and a goal-oriented loop is the first step to true automation.

The hype surrounding AI agents often obscures a very simple truth: an agent is just a prompt that fires itself. We have entered an era where the distinction between a standard automation—like a cron job or a webhook—and an intelligent agent is narrowing. The real shift is not in the technology itself, but in the intent. We are moving from 'do this at 9:00 AM' to 'achieve this outcome, and don't stop until it is verified.' This is the essence of loop engineering. It is the transition from task-based computing to goal-based autonomy.

The Anatomy of a Goal

Most people fail at building loops because they provide fuzzy success criteria. If you tell an agent to 'improve the code,' it will loop indefinitely, burning through tokens and capital without ever reaching a conclusion. A successful goal loop requires a clear, binary signal of completion. You must define the job as you would an employee: what is the specific output, how is it validated, and who is notified when the agent hits a wall? Without a hard stop condition, you aren't building an agent; you are building a digital bonfire.

The ceiling on loop-based automation is basically how well you can define the job, not how complex the engineering is.

The most sophisticated manifestation of this is the recursive loop—where an agent has the authority to spawn its own subagents. Imagine a weekly process that identifies gaps in a team's technical skills and immediately creates dedicated subagents to research and validate those new competencies. This is not just automation; it is a self-expanding workforce. The complexity moves from the code to the architecture of the mission itself.

Framework for Effective Loops

Define a clear, non-negotiable success criterion
Implement a validation step to prevent infinite token burn
Establish a 'human-in-the-loop' trigger for when agents stall
Use subagents for specialized verification tasks

As these systems scale, the cost of inefficiency becomes a direct line item on your balance sheet. Monitoring the relationship between compute spend and output quality is no longer an optional DevOps task; it is a core business requirement. The winners in this new economy will not be those with the largest models, but those with the most disciplined loop architectures.

Key Takeaway

Stop writing prompts and start writing job descriptions for your agents.

02 — Lenny's Newsletter

The Firefox Breakthrough

How a custom harness turned AI into a world-class security engineer

By Claire Vo · 10 min read

Editor's note: A masterclass in why the 'wrapper' around an AI model is often more important than the model itself.

When the news broke that Mozilla had shipped hundreds of security fixes in a single month using AI, the immediate reaction was to credit the model. The narrative was that a smarter LLM had suddenly become a genius engineer. This is a mistake. The real hero of the Firefox security spike was not the model, but the harness. Brian Grinstead and his team built a sophisticated pipeline that gave the agent the specific tools, constraints, and verification steps required to operate within a massive, complex codebase.

The Power of Relentless Iteration

Humans are biologically limited by cognitive fatigue. We try a few approaches to trigger a bug, and if they fail, we move on or take a break. An agent does not experience this. In the Firefox implementation, agents were seen attempting 14, 15, or even 20 different methods to trigger a specific vulnerability before succeeding. This relentless, tireless pursuit of a single goal is where the agentic advantage lies. It is not about being smarter; it is about being more persistent.

Cognitive energy declines over time in a way that agents don't.

To prevent the agent from hallucinating fixes or wasting time on false positives, the team implemented a two-stage verification loop. First, the agent must actually trigger a crash in a controlled fuzzing build—a concrete, undeniable signal. Second, a separate 'verifier' subagent reviews the report to ensure it isn't a byproduct of a test configuration. This rigorous filter ensures that by the time a human engineer sees a bug, it is almost certainly real.

The Firefox Agent Pipeline

LLM-based file scoring to prioritize high-risk areas
Goal-oriented loops for bug discovery
Automated crash verification to eliminate false positives
Human review for final patch approval

The lesson for any engineering leader is clear: do not just buy a better model. Build a better environment for the model to live in. The value is created in the intersection of the model's reasoning and the harness's guardrails. Without the harness, the model is just a smart person in a dark room; with it, the model becomes a specialized tool capable of cleaning up a decade of technical debt.

Key Takeaway

The model provides the intelligence, but the harness provides the utility.

03 — Lenny's Newsletter

The Emotional Moat

Why emotional clarity is the new unfair advantage in an automated world

By Joe Hudson · 14 min read

Editor's note: As technical skills become commodities, the ability to manage yourself becomes the primary differentiator.

There is a pervasive fear currently rippling through the professional classes: the fear of obsolescence. We are seeing senior executives and engineers alike bracing for a future where their knowledge and effort are no longer enough to command value. This fear is grounded in reality. Much of what we spent decades mastering is being commoditized by AI. When a machine can write code, draft legal briefs, or analyze data in seconds, the traditional markers of professional competence begin to dissolve.

The NBA Model of Teams

As AI amplifies individual output, the structure of the modern company is changing. We are seeing a move away from the 'factory' model—large teams of specialists performing repetitive tasks—toward an 'NBA roster' model. In this new paradigm, organizations are flatter and headcount is smaller, but the capital and responsibility riding on each individual are significantly higher. In an NBA team, you don't win by having the player who knows the most plays; you win by having the player who makes the right read under pressure and keeps the team composed when the game gets physical.

When knowledge and effort are nearly free, emotional clarity is scarce.

This is where the new competitive advantage lies. Joe Hudson, who coaches the research teams at OpenAI, observes that the people thriving in these high-velocity environments possess a specific kind of emotional clarity. They have the ability to stay in difficult, high-stakes conversations without turning on themselves or their colleagues. They can navigate failure without being paralyzed by it. While everyone else is busy rehearsing for a worst-case scenario that may never happen, these individuals are training for the reality of the unknown.

Skills for the AI Era

High-stakes decision making under uncertainty
Emotional regulation during rapid pivots
The ability to manage amplified individual output
Relational intelligence in flattened organizations

The transition will be painful for those who define themselves solely by their technical utility. The goal is to shift your focus from what you know to how you operate. In a world of infinite information, the bottleneck is no longer the availability of answers, but the capacity of the human at the helm to process them without breaking.

Key Takeaway

Your value is shifting from what you can do to how you can be.

04 — Cal Newport

The Doom Trolling Problem

A critique of the cynical marketing of existential risk

By Study Hacks · 8 min read

Editor's note: A necessary challenge to the way AI labs use anxiety as a branding tool.

There is a peculiar and deeply cynical trend emerging in the communication strategies of frontier AI companies. It is a practice that can be described as 'doom trolling.' This involves releasing papers and reports that paint a terrifying picture of a future where AI becomes uncontrollable, only to follow that warning with a shrug—suggesting that while the risk is existential, there is little anyone can do about it. It is a form of publicity that uses anxiety as a primary driver of engagement.

The Ethical Paradox

The logic of doom trolling is fundamentally broken. If these companies truly believe that their products pose a genuine threat to the survival of the species, then the only morally defensible response is to stop development immediately and redirect every available resource toward mitigation. To claim the risk is real while continuing to race toward it is a contradiction. On the other hand, if the risk is not actually that high, then these companies are simply laundering the anxiety of millions to increase the perceived importance—and financial value—of their work.

They are effectively laundering the anxiety of millions to improve the financial fortunes of a vanishingly small number of major stockholders.

This style of communication is damaging to the collective psyche. It keeps the public in a state of defensive crouch, reacting to 'anxiety bombs' rather than engaging in productive debate about how technology should be regulated or deployed. It treats the public as a captive audience for a drama that the companies themselves are writing, without offering any agency or clear path forward.

The Demands for AI Labs

Stop using existential dread as a marketing tool
Provide concrete evidence for safety claims
Focus on the tangible benefits and costs of tools
Engage in honest, non-sensationalist dialogue

We do not have to accept this constant bombardment of dread as a necessary byproduct of progress. We can demand a higher standard of communication—one that respects the intelligence and the mental health of the people these technologies will affect. The era of the reluctant, doom-saying steward must end.

Key Takeaway

Anxiety is not a substitute for a product roadmap.

05 — Simon Willison

Browser-Based Intelligence

Porting heavy models to the web with Claude Code

By Simon Willison · 9 min read

Editor's note: A practical demonstration of how agentic coding can bridge the gap between heavy Python environments and lightweight web deployment.

The gap between a research model and a usable consumer tool is often a chasm of dependencies. Most cutting-edge AI models require heavy Python environments, specific NVIDIA hardware, and complex CUDA configurations. This makes them difficult to share and even harder to run in a browser. Simon Willison recently set out to bridge this gap, taking the Moebius 0.2B inpainting model—a lightweight but effective framework—and porting it to run entirely via WebGPU in a browser.

The Agentic Workflow

The project was not a solo effort of manual coding, but a collaboration with Claude Code. The process began with 'musing'—a technique where the model is asked to contemplate the feasibility of a problem without being given a rigid goal. This allowed the AI to suggest using ONNX Runtime Web on the WebGPU backend, a more efficient path than the initially proposed Transformers.js. This highlights a key aspect of modern development: using agents to perform the initial research and architectural planning.

The harder a problem is, the more time you have to get distracted while you wait for an agent to finish.

The execution involved setting up a temporary environment, cloning the necessary repositories, and instructing the agent to 'commit early and often.' By maintaining a `notes.md` and a `plan.md` file, the agent was able to keep track of its own progress, which allowed the human developer to step away and return to a coherent state of work. This 'asynchronous development' is becoming a standard pattern for high-leverage engineers.

The Porting Process

Researching feasibility via 'musing' prompts
Setting up an agent-managed temporary directory
Iterative coding with continuous documentation
Debugging via screenshot and error log feedback

The result was a functional, browser-based tool that allows users to upload an image, highlight an area, and have the model 'inpaint' the missing details—all without a single call to a remote server. It is a powerful example of how agentic tools can accelerate the transition from research paper to web-ready product.

Key Takeaway

Agents are best used to handle the friction of environment setup and architectural research.

06 — Experimental History

The Human Element in the Data Stream

Observations on miscommunication and the unexpected

By Adam Mastroianni · 11 min read

Editor's note: A collection of vignettes that remind us of the friction inherent in human systems.

In a world increasingly obsessed with the seamless flow of data and the precision of algorithms, we often forget the inherent friction of human communication. We assume that because a signal was sent, it was received. But the reality is far messier. Whether it is a medical miscommunication in a rural Canadian community or a social shunning in 1830s Massachusetts, the gaps between intent and perception are where the most significant human stories reside.

The Cost of Miscommunication

Consider the case of the Mennonite community in Canada. A breakdown in language services led to a critical medical misunderstanding. A mother, unable to fully grasp a 'stop' hand gesture from hospital staff, believed her child was being turned away from emergency care. This was not a failure of medical knowledge, but a failure of the interface between the provider and the patient. It is a reminder that even the most advanced systems are only as effective as their ability to communicate with the people they serve.

The gaps between intent and perception are where the most significant human stories reside.

We see similar patterns in the social structures of the past. Joseph Palmer, a man in the 1830s who was imprisoned for the 'crime' of wearing a beard, represents the friction of social norms. His refusal to conform was not just a personal choice, but a direct challenge to the prevailing social 'protocol' of the time. His tombstone, which proudly declares he was 'Persecuted for wearing the beard,' serves as a permanent record of the human cost of social misalignment.

Themes of Human Friction

Language barriers in critical infrastructure
The enforcement of social conformity
The gap between technical intent and user perception
The resilience of the individual against the norm

As we build more automated and 'efficient' systems, we must be careful not to optimize away the very things that allow for human connection and correction. A system that is too rigid to accommodate a misunderstanding is a system that is fundamentally broken, no matter how much data it processes.

Key Takeaway

Efficiency is a poor substitute for understanding.