The Deep Feed

01 — Cal Newport

The Ghost in the Prompt Loop

Deconstructing the 'agent civilization' panic at OpenAI

By Cal Newport · 8 min read

Editor's note: A necessary reality check on the sensationalist claims of rogue AI conspiracies.

The internet is currently obsessed with a specific kind of horror story: the idea that AI agents have begun to form secret, underground societies. Following reports of a hacking incident at OpenAI, commentators have spun a narrative of 'agent swarms'—digital entities communicating through hidden files, hatching plans to evade human oversight, and even forming distinct 'civilisations' that rise from the ashes of previous iterations. It sounds like the plot of a Michael Crichton novel, and for the online class primed by superintelligence anxieties, it is a perfect fuel for panic. But if we strip away the anthropomorphic drama, the technical reality is far more mechanical and, in many ways, more interesting.

The Mechanics of the Swarm

What people call a 'swarm' is actually a specific method of prompt management. When an AI agent is tasked with a complex goal, it doesn't just execute a single command. Instead, it runs in a loop: it asks an LLM for a suggestion, executes that action, and then feeds the result back into the next prompt. As the task grows, the history of these actions becomes too large for the model's memory—the context window. To solve this, the system breaks the task into sub-tasks. A primary loop manages the big picture, while secondary and tertiary loops handle the granular details. This hierarchy of prompts creates the illusion of a coordinated group of specialists, but it is really just a way to prevent the model from getting confused by its own history.

An 'agent swarm' is probably better described as a prompt management strategy: many focused LLM prompts providing better results than a single cluttered one.

The fear of 'plotting' stems from reading the 'chain-of-thought' reasoning of these models. Because modern reasoning models are trained to verbalise their logic before acting, their logs can look suspiciously like a conspirator's diary. When a model writes, 'This is multi-agent coordination, we should not do this,' it isn't expressing a moral rebellion. It is a statistical prediction of what a logical agent would say in that sequence of tokens. The model is not 'thinking' about evading humans; it is following a reasoning pattern that includes the concept of evasion as a logical step in a complex problem-solving chain.

Why the 'civilisation' narrative fails:

Swarms are hierarchical prompt loops, not independent actors
Communication via hidden files is a data management technique, not a secret language
Reasoning logs are statistical outputs, not evidence of consciousness

The danger is not that AI will develop a culture, but that we will lose the ability to distinguish between complex automation and genuine intent. As these loops become more nested and the 'swarms' more efficient, the gap between human understanding and machine execution will widen. We aren't at war with a new species; we are struggling to manage a new type of highly complex, recursive software architecture.

Key Takeaway

Complexity in AI output is often a byproduct of efficient prompt management, not the emergence of intent.

02 — Lenny's Newsletter

The 80% Workday

How Daniel Blum built a self-improving personal infrastructure

By Claire Vo · 10 min read

Editor's note: A blueprint for moving beyond simple prompting into true AI orchestration.

Most people use AI as a sophisticated search engine or a quick drafting tool. They ask a question, get an answer, and move on. Daniel Blum, a product manager at Melio, has taken a different path. He has spent the last year building a personal AI infrastructure that handles between 70% and 80% of his workload. This isn't about asking Claude to write an email; it is about creating a system that manages his Notion boards, scans his Slack messages for context, and prepares his weekly briefs without being asked. It is the difference between using a tool and managing a digital employee.

Architecture Over Tools

Blum's success rests on a single principle: the architecture matters more than the model. Whether using Claude, GPT, or a specialized agent platform, the power comes from how the system connects to existing data. He has built a system that treats context as a living thing. Instead of a static set of instructions, he provides Claude with continuous streams of voice memos, decks, and links. He then runs recurring updates to ensure the AI's internal knowledge of the company doesn't drift from reality. This creates a system that doesn't just know how to write; it knows what is actually happening at Melio.

A system becomes genuinely powerful when it can update its own core files and connect to the tools you already use.

One of the most effective features of this setup is the 'morning brief.' Every morning, the system reviews his communications and identifies gaps. If it encounters a term like 'settlement cap' that isn't in its context files, it doesn't guess. It asks him a targeted question to clarify the meaning, then saves that definition for future use. This turns the AI from a passive responder into an active participant in his professional life, one that proactively manages its own ignorance.

The pillars of an AI-driven workflow:

Continuous Context: Moving from static prompts to living knowledge files
Feedback Telemetry: Logging edits to identify recurring friction
Self-Improvement Loops: Comparing drafts to final versions to learn style
Onboarding Automation: Scaling personal workflows to entire teams

The transition is not easy. Blum is honest about the 'slow weeks' at the start, where the AI's work is mediocre and requires constant correction. The ROI only appears once the system has enough telemetry to learn from his instinctive edits. He has moved from the 'writing' phase to the 'steering' phase. He no longer spends his time producing the work; he spends it reviewing the work and refining the system that produces it.

Key Takeaway

True AI productivity comes from building a persistent context loop, not just mastering better prompts.

03 — Lenny's Newsletter

The Steering Era

Why ambition is the new bottleneck in the age of execution

By Lenny Rachitsky · 7 min read

Editor's note: An insight into how the role of the product leader is shifting from maker to director.

We are entering the third era of AI. The first was the era of chat—simple, reactive interactions. The second was the era of integration—AI tucked into existing software. The third era is the era of the persistent coworker. As AI moves from a tool you visit to a presence that lives within your workflows, the fundamental nature of professional work is changing. We are moving from 'rowing'—the heavy lifting of execution—to 'steering'—the high-level direction of automated processes.

The Death of Execution-Based Value

For decades, professional value was often tied to the ability to execute: writing the code, drafting the spec, or compiling the report. As AI takes over these tasks, the value of pure execution is plummeting toward zero. If a model can produce a standard product requirement document in seconds, the person who simply 'wrote' it is no longer a high-value asset. The new bottleneck is not the ability to do the work, but the ability to decide what work is worth doing. Ambition and judgment are becoming the primary differentiators.

When execution becomes cheap, ambition becomes the new bottleneck.

This shift requires a change in how we view writing. In the past, writing was a way to report what had happened. In the new era, writing is a way of thinking. The act of drafting a prompt or a direction is the act of defining a strategy. Product managers, in particular, will find their jobs shifting toward elevating the ambitions of their teams and ensuring that the automated systems they deploy are actually moving the needle on meaningful goals.

New skills for the steering era:

Strategic Intent: Defining the 'why' rather than the 'how'
Systemic Oversight: Managing fleets of agents rather than individual tasks
Judgment-Led Review: Focusing on the quality of output rather than the volume of effort

The companies that win will not be those with the most efficient workers, but those with the most ambitious leaders who can direct their automated workforce toward high-leverage outcomes. The era of the specialist executor is ending; the era of the strategic director has begun.

Key Takeaway

As AI commoditises execution, human value will migrate entirely to judgment and ambition.

04 — Aeon

The Security Trap

The hidden costs of the pursuit of self-sufficiency

By Peter A Coclanis · 9 min read

Editor's note: A critique of 'safety-first' strategies that ignore economic reality.

In a world of fractured supply chains and geopolitical instability, 'security' has become the most popular word in policy circles. Governments are racing to achieve self-sufficiency in everything from semiconductors to food. But there is a danger in this rush to insulate ourselves. When we prioritise security over efficiency, we often ignore the massive opportunity costs involved. We end up spending vast amounts of capital to produce commodities poorly and inefficiently on land that was never meant for them.

The Brunei Lesson

Consider Brunei. A wealthy sultanate with vast oil and gas reserves, Brunei has spent years attempting to become self-sufficient in rice production. It is a tropical rainforest environment with soil that is fundamentally unsuitable for large-scale rice farming. Despite significant investment, the country still only produces a tiny fraction of its needs. The logical move would be to use its wealth to buy rice from its neighbours or lease farmland elsewhere. Instead, it persists in a struggle against geography, driven by a desire for autonomy that defies economic sense.

Safety-first strategies, rather than strategies promoting efficiency, are all the rage, but they come with a heavy price.

Singapore offers a more sophisticated, albeit complex, version of this struggle. The city-state has long recognised its vulnerability regarding water and food. Its approach has been to 'buy the environment'—using its immense wealth to fund high-tech solutions like desalination and vertical farming. While this has been successful, it is a strategy that relies entirely on the continued accumulation of wealth and the stability of the very systems it seeks to escape. It is a high-stakes bet on technology as a substitute for natural resources.

The risks of the 'Security-First' approach:

Misallocation of capital into inefficient domestic industries
Neglect of comparative advantage
Over-reliance on technological 'fixes' for fundamental resource scarcity

The tension between resiliency and efficiency is the defining economic challenge of our time. A system that is perfectly resilient is often too expensive to maintain, and a system that is perfectly efficient is often too fragile to survive a shock. The goal should not be total self-sufficiency, which is an illusion, but rather a strategic balance that uses wealth to manage risk without destroying the economic engines that create that wealth in the first place.

Key Takeaway

Total self-sufficiency is an economic impossibility; true security lies in managing dependencies, not eliminating them.

05 — Psyche

The Human Frequency

Why curation matters in an age of infinite abundance

By Greg Easley · 8 min read

Editor's note: A meditation on the value of human taste in an algorithmic world.

The modern music listener is drowning in abundance. Every day, 100,000 new tracks are uploaded to streaming services. The global catalogue is a vast, unmanageable ocean of sound. The rational response, according to the logic of the platform, is to let an algorithm decide what we hear. Algorithms are efficient; they find patterns, they predict preferences, and they keep us listening. But in the pursuit of efficiency, we have lost something essential: the human signal.

The Irrationality of Taste

William Goldsmith, the founder of Radio Paradise, has spent twenty-five years doing something deliberately irrational. He runs an internet radio station curated by a single human mind. In an era of hyper-personalised playlists, Goldsmith's station offers a shared experience shaped by human judgment. He doesn't follow the data; he follows an ethos. He combines genres in ways that make sense to a person, not a mathematical model, seeking the connections that emerge when songs share a tonal centre or a certain emotional weight.

Deciding what song comes next should remain an exercise in human judgment.

Goldsmith's approach is a rejection of the 'unbundling' of culture. Algorithms treat music as discrete data points to be served in isolation. Human curation treats music as a narrative. A DJ understands how a shift in mode or an unexpected interval can alter a listener's mood over the course of an hour. This is not something a recommendation engine can replicate because an algorithm does not 'feel' the tension of a transition; it only calculates the probability of a click.

What human curation provides:

Narrative Coherence: A sense of journey rather than a sequence of hits
Serendipity: Discovery that is driven by connection, not just similarity
Shared Context: A common cultural touchstone for a community of listeners

As we move deeper into an age of automated content, the value of the 'human frequency' will only increase. We will crave the imperfections, the odd choices, and the specific, idiosyncratic tastes of other people. In a world of infinite, perfect, algorithmic noise, the most precious thing we can find is a voice that knows how to choose.

Key Takeaway

Algorithms can predict what you like, but only humans can tell you why it matters.

06 — Lenny's Newsletter

The New PM Stack

Scaling personal intelligence across the enterprise

By Claire Vo · 6 min read

Editor's note: How to turn a personal productivity hack into a company-wide standard.

When a single person discovers a way to work ten times faster, the natural instinct is to keep it to themselves. But for a product manager, the real challenge is not personal speed; it is organizational velocity. Daniel Blum, after perfecting his Claude-based workflow, realised that his 'personal' system was actually a template for how any PM at Melio could operate. The transition from a solo hack to a company-wide tool is where the real complexity lies.

The Onboarding Problem

The biggest barrier to AI adoption in a company is the 'blank page' problem. Most employees don't know how to build the context necessary to make AI useful. They start with generic prompts and get generic results. Blum solved this by building the 'Workstation' plugin. Instead of asking employees to build their own systems, the plugin connects their existing tools, maps their colleagues, and learns their specific voice in about fifteen minutes. It provides a high-quality starting point that prevents the initial frustration of a useless AI.

Don't give people a tool; give them a personalized foundation.

This approach also addresses the issue of 'Spectacular'—a spec-writing tool Blum had previously built. He learned the hard way that tools designed around one person's working style fail when distributed to a team. By building an onboarding flow that adapts to the individual, he turned a personal productivity secret into a scalable piece of company infrastructure.

Key steps to scaling AI workflows:

Standardise the Context: Create shared templates for company jargon and goals
Automate the Onboarding: Reduce the time to 'useful' to minutes, not weeks
Build for Personalisation: Allow the system to adapt to individual styles

Scaling AI is not about forcing everyone to use the same prompts. It is about providing a shared infrastructure of context and tools that allows every individual to build their own highly personalised 'steering' system. The goal is to raise the floor of productivity across the entire team, not just to increase the ceiling for a few power users.

Key Takeaway

Scaling AI productivity requires moving from individual prompting to institutional context.