The Deep Feed

01 — Lenny's Newsletter

The Agent Harness: Moving Beyond the Chat Box

Why the future of AI productivity lies in specialized code, not better prompting

By Lenny Rachitsky · 12 min read

Editor's note: The era of 'talking' to AI is ending; the era of building around it has begun.

The current obsession with the 'perfect prompt' is a distraction. We treat large language models like magic lamps, rubbing them and hoping for a specific wish, but this approach is fundamentally unscalable for professional work. The real breakthrough isn't in how we talk to the model, but in how we wrap the model in code. This is the concept of the 'harness'. A harness is a layer of deterministic logic that sits around a non-deterministic AI agent, providing the constraints, tools, and structure necessary to turn a chatty assistant into a reliable worker. When you give an agent a general-purpose interface, you are essentially giving a highly skilled carpenter a blunt hammer and asking them to 'figure it out'. A harness, by contrast, provides the specific workbench, the exact measurements, and the safety guards required to produce a repeatable result.

The Death of General-Purpose Agents

General-purpose tools like Claude Code or Codex are impressive, but they suffer from a lack of opinion. They wander through data, making guesses that might be right but are rarely consistent. For a professional workflow—such as triaging engineering bugs—consistency is more valuable than raw intelligence. By building a custom harness, you can encode permissions directly into the system. You no longer need to remind the agent to 'only investigate and not write code'; the harness simply does not provide the functions required to write code. This specificity makes the agent faster, cheaper, and significantly more reliable. It moves the intelligence from the prompt to the architecture.

Giving a constrained agent a specific harness gets more consistent output than giving a powerful agent an open prompt.

Furthermore, a harness allows for multi-model routing. A single chat interface locks you into one provider's logic. A custom harness allows you to use a lightweight, fast model for initial data gathering, a heavy-duty model for reasoning, and a specialized model for final formatting. You are no longer a consumer of an AI product; you are the architect of an AI system. This shift from 'using' to 'orchestrating' is the defining transition for anyone looking to maintain an edge in an automated economy.

How to identify a harness-ready task

The workflow has defined steps and tools (deterministic).
The outcome requires reasoning or synthesis (non-deterministic).
The task is repeatable and requires a consistent output format.
The agent needs specific data access that shouldn't be granted to a general tool.

Ultimately, the goal is to create structured artifacts. A one-off chat response is a fleeting moment of utility. A harness that outputs a task log, a technical brief, and an HTML summary creates a permanent, scannable record for a team. This turns AI from a novelty into a piece of infrastructure. We are moving away from the era of 'AI as a companion' and into the era of 'AI as a component'.

Key Takeaway

Stop trying to prompt your way to excellence; start building the code that constrains your agents into being useful.

02 — Lenny's Newsletter

The 24/7 Software Factory

The economics of unlimited local inference

By Claire Vo · 10 min read

Editor's note: Why the smartest builders are moving their intelligence off the cloud and onto their own desks.

There is a common misconception that the path to AI dominance is through a $20 monthly subscription to a cloud provider. For a casual user, this is true. For a builder running continuous loops of code review, security scanning, and social monitoring, it is a financial trap. Alex Finn has demonstrated a different path: the local AI fleet. By investing in high-end hardware—Mac Studios, Nvidia DGX units, and custom RTX builds—he has achieved something cloud users cannot: unlimited inference. When the cost of a single reasoning step is effectively zero once the hardware is paid for, the entire math of what is worth automating changes.

Hardware as Strategy

Running a local fleet requires more than just a fast GPU; it requires an understanding of how different hardware tiers serve different cognitive tasks. A Mac Studio with massive unified memory is not for speed; it is for scale, allowing the running of massive models that would otherwise require a server farm. An RTX 5090 is for velocity, handling the rapid-fire iterations of a build-and-review loop. The strategy is to match the model's requirements to the machine's specific strengths. This is not about having the most powerful computer; it is about having the right distribution of compute.

The case for local AI isn’t ROI; it’s unlimited inference.

This setup enables 'ambient AI'—a state where agents run in the background, constantly monitoring, testing, and building without human intervention. Using tools like Tailscale to connect these machines into a single network, a builder can manage a fleet of agents that act as a continuous software factory. One agent might be scanning for vulnerabilities, while another is writing documentation, and a third is running browser tests. This is not a tool you use; it is a workforce you manage.

The Local AI Stack

Connectivity: Tailscale for multi-machine networking.
Orchestration: OpenClaw or Hermes for agent management.
Models: GLM for high intelligence, Qwen for speed, and Ornith for specialized tasks.
Testing: Playwright for automated browser-based verification.

The barrier to entry is no longer just intelligence, but the ability to manage physical and digital infrastructure. As cloud costs scale linearly with usage, the local builder scales their capability with their hardware investment. It is a shift from an OpEx-heavy model to a CapEx-heavy model, and for those running 24/7 operations, the transition is not just beneficial—it is inevitable.

Key Takeaway

If your AI workflow requires constant, repetitive reasoning, stop renting intelligence and start owning the hardware.

03 — Lenny's Newsletter

The Great Tech Split

Sentiment and survival in the age of automation

By Lenny Rachitsky · 15 min read

Editor's note: The annual sentiment survey reveals a workforce dividing into two distinct camps.

The tech industry is no longer a monolith of optimistic progress. According to the latest Tech Worker Sentiment Survey, the workforce has split almost exactly in half. On one side are the 'Energized'—those who have integrated AI into their workflows to multiply their output and find new creative freedoms. On the other are the 'Resentful' and 'Disoriented'—workers who feel the ground shifting beneath them, watching their specialized skills be commodified in real-time. This isn't just a difference in opinion; it is a divergence in career trajectory.

The Burnout Paradox

While AI promises to reduce drudgery, burnout has actually increased by 11 points in a single year. This seems contradictory, but the reason is clear: the baseline for 'acceptable productivity' has moved. When a tool allows you to write code ten times faster, the expectation is not that you work ten times less, but that you produce ten times more. The velocity of the industry has accelerated, and the human nervous system is struggling to keep pace. The pressure is no longer just to be good, but to be infinitely productive.

The primary fear in tech today is not job loss to AI, but the loss of agency and the erosion of craft.

The survey also highlights a startling trend: almost no one in tech would recommend their job to a newcomer today. The industry has moved from a place of building tangible things to a place of managing automated flows. For many, the sense of mastery—the 'flow state' that comes from solving a difficult problem—is being replaced by the 'supervisory state', where one simply checks the work of a machine. This shift in the nature of work is driving a crisis of meaning.

The Four Archetypes of the AI Era

The Energized: Using AI to expand the boundaries of what they can build.
The Conflicted: Seeing the utility but fearing the long-term implications.
The Disoriented: Struggling to find where their value lies in an automated world.
The Resentful: Feeling that the industry's soul is being traded for efficiency.

For managers, the takeaway is stark. Employee well-being is no longer about perks or remote work; it is about helping workers find a way to remain relevant and engaged in a high-velocity environment. The winners in this new era will be those who can bridge the gap between human intuition and machine speed without losing their sense of purpose.

Key Takeaway

AI is not just changing what we do; it is changing how we feel about doing it.

04 — Cal Newport

The Literacy Crisis

Why the decline of deep reading is a cognitive catastrophe

By Study Hacks · 8 min read

Editor's note: We are losing the ability to think deeply because we are losing the ability to read deeply.

We are witnessing a quiet, systemic erosion of the human intellect. The data is difficult to ignore: reading for pleasure has plummeted, and the ability of even high school seniors to interpret complex text is failing. This is not merely a change in how we consume information; it is a fundamental change in how we process thought. Reading is not a passive act of receiving data; it is a cognitive exercise that shapes the very structure of our consciousness. When we abandon long-form text for the fragmented, rapid-fire consumption of digital snippets, we are effectively de-training our brains.

The Technology of the Mind

Literacy is more than a skill; it is the technology that enabled the modern world. Concepts like logic, individual rights, and scientific progress are all conceptual children of the written word. The ability to hold a complex, multi-page argument in one's mind is what allows for deep reasoning and long-term planning. As we shift toward a 'skimming' culture, we lose the capacity for the kind of sustained attention required to solve the world's most difficult problems. We are trading depth for breadth, and the cost is our ability to think for ourselves.

Literacy is not simply a technology, but the very architecture of the cognitive world we take for granted.

The digital assault on literacy is often framed as a matter of distraction, but it is deeper than that. The architecture of the internet—the infinite scroll, the notification, the algorithm—is designed to keep us in a state of shallow engagement. This state is the antithesis of reading. Reading requires a voluntary surrender to a single line of thought; the internet requires a constant, reactive scanning of the environment. We are becoming expert scanners but amateur thinkers.

Signs of Cognitive Decline

Inability to paraphrase a multi-page text.
A preference for short-form video over written explanation.
Difficulty maintaining focus on a single idea for more than a few minutes.
The loss of the ability to engage with complex, non-linear arguments.

To resist this, we must treat cognitive fitness with the same seriousness we treat physical health. We need routines that prioritize deep, slow information consumption. If we do not, we risk entering a new kind of dark age—one where we have access to all the world's information but lack the mental tools to understand any of it.

Key Takeaway

Deep reading is the gym for your intellect; without it, your capacity for complex thought will atrophy.

05 — Stratechery

The Apple-OpenAI Conflict

The real battle for the interface of intelligence

By Stratechery · 7 min read

Editor's note: A lawsuit over trade secrets is a smokescreen for a much larger fight over who controls the user experience.

Apple's recent legal action against OpenAI may look like a standard corporate dispute over trade secrets, but it is actually a symptom of a much larger existential crisis. For decades, Apple has maintained its dominance by controlling the interface between the user and the machine. They own the hardware, the operating system, and the ecosystem. But AI is a new kind of layer—one that doesn't care about operating systems. When a user interacts with an LLM, they are interacting with a brain that exists outside of Apple's walled garden. This threatens the very foundation of Apple's business model.

The Threat of the Unbundled Interface

The danger for Apple is not that OpenAI will steal their code, but that OpenAI will make the iPhone irrelevant. If the primary way a human interacts with their technology is through a conversational agent, the underlying device becomes a mere commodity—a vessel for the connection. If the 'intelligence' lives in the cloud and the 'interface' is a voice or a text box, the value shifts from the device manufacturer to the model provider. Apple is fighting to ensure that the AI remains a feature of the OS, rather than a replacement for it.

Apple's real problem is not intellectual property theft; it is the unbundling of the user experience.

This conflict highlights the tension between the 'platform' model and the 'intelligence' model. Platforms win by creating friction-less ecosystems that keep users locked in. Intelligence models win by providing the most capable reasoning, regardless of where it is accessed. As AI becomes more capable, the friction that Apple relies on to maintain its moat begins to dissolve. The more helpful an agent becomes, the less a user cares about which brand of phone they are holding while they talk to it.

The Battlefronts of AI Dominance

The Model: Who provides the core reasoning?
The Interface: Who controls the primary point of interaction?
The Data: Who owns the personal context used to refine the agent?
The Hardware: Does the device still matter if the intelligence is remote?

Apple's strategy will likely be to integrate AI so deeply into the OS that it becomes impossible to separate the two. They want to move from being a provider of tools to being the provider of the very cognitive layer through which we experience the world. The lawsuit is a defensive move in a war that is being fought not in the courts, but in the pocket of every consumer.

Key Takeaway

In the age of AI, the winner won't be the one who builds the best device, but the one who controls the most important conversation.

06 — The Marginalian

Chronodiversity: Living in Multiple Times

Beyond the clock: finding meaning in the elasticity of time

By Maria Popova · 9 min read

Editor's note: Time is not a resource to be spent, but a dimension to be inhabited.

We are taught to view time as a linear resource—a commodity to be managed, saved, or wasted. This view is a trap. It reduces the richness of existence to a series of transactions. To truly live is to recognize 'chronodiversity': the idea that time is not a single, uniform stream, but a collection of different ways of being. There is the chronological time of the clock (chronos) and the qualitative time of the moment (kairos). There is the rapid, frantic time of the insect and the slow, enduring time of the star. When we stop trying to master time and start learning to inhabit its different modes, we find a way to be more fully alive.

The Elasticity of Experience

Time is famously elastic. It stretches during moments of fear and accelerates during periods of novelty-starved routine. This is not a failure of our perception, but a fundamental property of how we experience existence. The 'fast time' of elderhood, where the lack of new experiences makes the years blur together, is the inverse of the 'slow time' of youth, where every sensation is sharp and expansive. Recognizing these shifts allows us to understand our own psychological states rather than fighting against them. We can learn to seek out the novelty that slows time down, or the stillness that allows us to catch our breath.

Time is not a resource we have for cashing in; it is a dynamic and unpredictable array of co-existing realities.

Living with 'timefulness' means acknowledging the relational nature of time. Your time is not the same as mine. In an unequal world, the ability to relax is a luxury that others cannot afford. By understanding that time is contingent and localized, we can move away from the rigid, industrial standards of productivity and toward a more human, rhythmic way of being. We can learn to honor the 'slow time' required for deep thought and the 'fast time' required for urgent action.

Ways to Inhabit Time

Chronos: Respecting the structure and rhythm of the clock.
Kairos: Seeking the qualitative, meaningful moments that exist outside of schedules.
Relational Time: Acknowledging how our time interacts with the time of others.
Deep Time: Connecting to the vast, geological, or cosmic scales of existence.

Ultimately, the goal is not to control time, but to be present within it. Whether we are experiencing the fleeting seconds of a heartbeat or the slow movement of the seasons, the value lies in our awareness. To live well is to navigate the tension between the transient and the eternal, finding meaning in the very fact that our time is both limited and infinitely varied.

Key Takeaway

Stop trying to manage your time and start learning how to inhabit its different rhythms.