The Deep Feed

01 — Stratechery

The Agentic Guard: When AI Protects the Machine

How persistent monitoring turned a security breach into a diagnostic triumph

By Stratechery · 8 min read

Editor's note: A practical look at the tension between Apple's walled garden and the necessity of autonomous agents.

A macOS vulnerability, tracked as CVE-2026-65400, recently allowed attackers to gain root access via a screen-sharing flaw. The goal was simple: plant a Monero crypto miner. Most users would have discovered the breach through a sluggish system or a suspicious bill. However, the author's experience suggests a different future. An autonomous agent, built to manage projects and track ideas, acted as a persistent sentry. It didn't just watch; it reacted. When the agent's scheduled monitoring tool restarted, it detected a deviation in system permissions. It unilaterally stopped executing commands and issued an urgent warning: the account could now run admin commands without a password. This wasn't a chatbot providing advice; it was a system identifying a breach in real-time.

The Conflict of Control

This incident exposes the growing rift between platform owners and agent developers. Apple’s recent stance on Full Disk Access reflects a defensive posture. They view apps that request deep system permissions as risks to user privacy. From Apple's perspective, an app with full access can scrape mail, messages, and browsing history. They are moving toward a model of 'very explicit user action' to gate these permissions. But for the power user, this creates a paradox. To build an agent capable of true protection—an agent that can monitor system states and respond to threats—the agent requires the very access Apple is trying to restrict. The more useful an agent becomes, the more it looks like a security threat to the operating system.

The more useful an agent becomes, the more it looks like a security threat to the operating system.

The author's response to the hack was not to retreat from AI, but to lean into it. Instead of manually hunting for the intruder, they used Claude to root out the malware, identify the exact four-second window of the exploit, and build a custom tool to prevent a recurrence. This is the shift from reactive security to proactive, agent-led defense. We are moving away from a world where we wait for a patch and toward a world where our software actively defends our digital perimeter. The risk is real, but the cost of not having an agent may soon be higher.

Agent Security Principles

Constrain capabilities to specific, necessary domains.
Implement scheduled restarts to clear state and detect anomalies.
Use persistent monitoring as an automated inbox for system health.
Treat agent warnings as high-priority diagnostic signals.

Ultimately, the battle for the desktop is a battle over who manages the state of the machine. Is it the OS manufacturer, or is it the user's personal agent? As agents become more capable of executing code and managing files, the traditional security models of 'permission-based access' will struggle to keep up with the speed of automated threats and automated defenses.

Key Takeaway

True digital security in the agentic era requires autonomous systems that can monitor and react faster than a human can react to a notification.

02 — Lenny's Newsletter

The Death of the Chatbot

Why the future of AI lies in decision engines, not conversation

By John Lindquist · 10 min read

Editor's note: A technical shift that changes how developers should approach AI integration.

The prevailing mental model for AI is the chatbot: a window where you type a question and receive a paragraph of text. This model is inefficient for most production tasks. The real value is emerging in a different category: the decision engine. Tools like Jev do not aim to converse; they aim to output structure. They take unstructured input—a voice command, a messy JSON file, a chess move—and return a classification, a score, or a function call. By stripping away the need for conversational fluff, these models become incredibly fast and cheap. When a decision costs a fraction of a cent and happens in milliseconds, the architectural possibilities change.

From Prompting to Logic

Developers often approach AI through 'prompt engineering,' trying to coax the right answer out of a large, expensive model. The new approach is closer to traditional programming. Instead of asking a model to 'think about this,' you use it as a high-speed if/else statement. You define the available paths—the routes through an application—and the model simply selects the correct one. This is routing. A single natural language input can identify the intended action and trigger the appropriate tool. This is how voice assistants finally become useful: not by talking back, but by executing the command instantly.

The right mental model for a decision engine is an if/else statement, not prompt engineering.

This efficiency enables a strategy of 'efficient inefficiency.' In traditional software, we try to narrow the search space to save compute. With cheap decision engines, we can afford to be brute-force. We can evaluate every possible move in a chess game or every route through a database because the cost of being wrong or being thorough is negligible. We can run layered classifications, where one cheap model's output feeds into another, increasing accuracy through repetition rather than through the sheer size of a single model.

When to use Decision Engines

Routing natural language to specific functions.
Classifying data into predefined categories.
Scoring or ranking unstructured inputs.
Triggering automated workflows based on intent.

However, this is not a total replacement for Large Language Models. Creative reasoning, open-ended brainstorming, and complex synthesis still require the heavy lifting of generative models. The most effective systems will be hybrids: using cheap, fast decision engines to handle the logic and routing, and reserving the expensive, high-reasoning models for the moments that actually require thought.

Key Takeaway

Stop building chatbots and start building decision engines that turn messy input into structured action.

03 — Aeon

The Public Self

How digital archives and social feeds reshape the human narrative

By Thomas Söderqvist · 12 min read

Editor's note: A philosophical investigation into how publicizing our lives changes our internal identity.

For most of history, the act of writing a memoir was a solitary, retrospective process. An author would spend years gathering papers, consulting friends, and refining a narrative before presenting a finished, coherent version of their life to the world. The public only saw the result—the polished, static account. Social media has destroyed this distance. By publishing life in sequential episodes, the author turns autobiography into a real-time dialogue. The narrative is no longer a closed loop; it is an open system, constantly being revised by the reactions and interpretations of a digital audience.

The Mirror of the Audience

This shift changes how we perceive ourselves. When we write in public, we are not just recording history; we are performing it. The feedback loop of a digital audience acts as a mirror, often reflecting aspects of our character that we were blind to in private. The author notes a moment where a reader pointed out that a traumatic story about his origins was being told with a 'deadpan' irony—a detail he hadn't consciously realised. The public eye doesn't just watch; it interprets, and in doing so, it forces the storyteller to confront the gap between their intention and their impact.

What happens to self-image when it is formed in public?

This raises a fundamental question about the coherence of the self. If our identity is being shaped by the continuous, fragmented input of others, can we ever truly claim a unified narrative? The traditional biography seeks to find meaning in a life's trajectory. But the digital life is a series of disconnected episodes, subject to the whims of engagement and the shifting perspectives of a crowd. We are moving from a model of 'life as a book' to 'life as a live feed,' where the sense of a stable, central 'I' is constantly under negotiation.

There is a certain danger in this transparency. When the boundary between the private thought and the public post dissolves, the space required for internal reflection shrinks. If every experience is immediately processed through the lens of how it will be received, the authenticity of the experience itself is compromised. We risk becoming curators of our own lives rather than participants in them, constantly editing our present to fit a narrative that the audience expects.

Key Takeaway

Publicly documenting your life turns your identity into a collaborative project, often at the expense of a coherent, private self.

04 — Psyche

The Resilience Portfolio

Redefining strength in the face of modern motherhood

By Sarah Emmerson · 7 min read

Editor's note: A psychological reframing of motherhood that moves away from individual deficiency.

Modern motherhood is often framed as a test of individual character. Women frequently find themselves comparing their struggles to an impossible blueprint—the 'strong' mother of previous generations or the curated, effortless versions seen on social media. This creates a sense of inherent deficiency: the belief that if they cannot cope, it is because they lack the 'resilience' that others seem to possess. This perspective is damaging because it treats resilience as a fixed trait—something you either have or you don't—rather than a dynamic state influenced by external factors.

Beyond Character Traits

Psychological research suggests a more useful model: the 'resilience portfolio.' Resilience is not a single muscle; it is a collection of assets across four distinct domains. These include meaning-making (purpose and hope), regulatory strengths (emotional management), interpersonal resources (social support), and environmental context (access to stable structures). When a mother struggles, it is rarely because she is 'weak.' It is more likely because her portfolio has been systematically dismantled by her circumstances.

Strength is not a character trait; it is a portfolio of assets that can be built and replenished.

For many new mothers, the transition to parenting strikes at the very heart of their meaning-making domain. Women who previously derived identity from professional achievement or personal autonomy often find those pillars temporarily removed. When the ability to 'get things done' is replaced by the repetitive, stationary demands of infant care, the sense of purpose can fracture. This is not a failure of will; it is a structural shift in how they interact with the world.

The Four Domains of Resilience

Meaning-making: Finding purpose and hope.
Regulatory strengths: Managing and reflecting on emotions.
Interpersonal resources: Access to reliable social support.
Environmental context: Stability in childcare, finances, and community.

To support mothers, we must stop asking them to be 'stronger' and start asking how we can help them rebuild their portfolios. This means addressing the environmental and interpersonal gaps—childcare costs, maternity pay, and social isolation—that make resilience impossible. Resilience is a collective responsibility, not an individual burden. If the environment is designed to deplete a person's resources, no amount of 'inner strength' will be sufficient.

Key Takeaway

Resilience is a combination of personal skills and external support; stop blaming mothers for failing in environments that deny them both.

05 — Lenny's Newsletter

The Infrastructure of Personal Software

How ChatGPT Sites is turning prototyping into production

By Kath Korevec · 9 min read

Editor's note: An inside look at how OpenAI employees are using their own tools to build internal infrastructure.

The distinction between a prototype and a production tool is traditionally defined by the complexity of the underlying architecture. A prototype is a fragile sketch; production software is a robust system with storage, authentication, and integration. ChatGPT Sites is attempting to collapse this distinction. By providing an infrastructure layer—including D1 storage, R2 buckets, and MCP plugin hosting—OpenAI is enabling users to build tools that aren't just for testing, but for actual work. This is the shift from 'chatting with an AI' to 'building with an AI'.

The Power of Connectors

The real utility of these sites lies in their ability to act as connective tissue between disparate data sources. Through Plugin Insights, a site can automatically personalise itself for every visitor. For example, an incident command center built by the OpenAI product team doesn't just show a static list of bugs. It uses Slack and Notion connectors to infer a user's team and display only the relevant runbooks and active incidents. The user doesn't have to write the logic for this; the system handles the authentication and data retrieval, allowing the user to focus on the utility of the tool.

Software can now be temporary and still worth building.

This leads to a new way of thinking about software development. We are entering an era of 'disposable software'—tools that are highly specialised for a specific task, a specific trip, or a specific project, and are discarded once the task is complete. Because the infrastructure is managed and the connectors are pre-built, the cost of creating a bespoke tool for a one-week use case is negligible. This allows for a level of granular automation that was previously too expensive or time-consuming to justify.

Key Features of ChatGPT Sites

Integrated storage (D1 and R2) for persistent data.
MCP plugin hosting for seamless tool integration.
Plugin Insights for automatic user personalisation.
Support for co-editing and collaborative development.

As these tools mature, the barrier to entry for building functional, data-driven applications will continue to fall. The goal is a world where anyone can describe a workflow and have a custom, integrated software interface appear to manage it. We are moving from a world of general-purpose apps to a world of infinitely many, highly specific, AI-generated tools.

Key Takeaway

The future of software is not in monolithic apps, but in the ability to instantly generate bespoke, data-connected tools for specific tasks.

06 — Lenny's Newsletter

The Agentic Shift

Navigating the transition from tools to teammates

By Tibo Sottiaux · 8 min read

Editor's note: A look at the strategic direction of OpenAI's consumer products.

We are witnessing a fundamental shift in the nature of human-computer interaction. For decades, software has been a collection of tools: a hammer, a calculator, a word processor. You pick them up, use them, and put them down. The new era, defined by platforms like OpenAI's Dots, is moving toward agents. An agent is not a tool; it is a teammate. It doesn't just wait for a command; it understands intent, manages workflows, and acts on your behalf across the digital landscape. This represents a transition from 'software as a utility' to 'software as an agentic partner'.

The End of the Manual Loop

Most current AI interactions are still trapped in a loop of manual prompting and manual verification. You ask, it answers, you check, you correct. The goal of agentic platforms is to break this loop. By moving toward 'graphs' and 'loops' of autonomous reasoning, agents can handle multi-step processes without constant human intervention. They can plan a trip, book the flights, and adjust the itinerary when a delay occurs, all while operating within the constraints and preferences you have set. The value shifts from the ability to generate text to the ability to execute complex, real-world actions.

Most actions on the internet will soon be taken by agents, not humans clicking buttons.

This shift necessitates a change in the skills required to work in the AI era. The ability to write perfect prompts is becoming less important than the ability to design robust workflows and supervise agentic behavior. We are moving from being 'operators' of software to being 'managers' of agents. This requires a different kind of oversight—one focused on setting boundaries, defining objectives, and managing the edge cases where an agent might deviate from its intended path.

Skills for the Agentic Era

Workflow design and orchestration.
Boundary setting and constraint management.
High-level objective definition.
Supervisory oversight of autonomous processes.

As these agents become more integrated into our lives, the tension between autonomy and control will become the central theme of technology. How much agency do we grant our digital teammates? How do we ensure they act in our best interest while still allowing them the freedom to be effective? The answer will define the next decade of human-machine collaboration.

Key Takeaway

The next era of computing is not about better tools, but about more capable agents that move from executing commands to managing intentions.