The Deep Feed

01 — Lenny's Newsletter

The Google I/O Delusion: Agents, Tools, and the Gap Between Hype and Utility

A breakdown of the 2026 announcements and why availability remains the industry's biggest failure.

By Claire Vo · 12 min read

Editor's note: Google is moving fast, but the distance between a demo and a working product is still a canyon.

Google I/O 2026 arrived with the usual fanfare, a bombardment of new models and creative tools designed to convince us that the era of the 'agent' is finally here. The headline act was the Gemini 3.5 model family, specifically the Flash variant, which aims to win the race on speed and coding efficiency. But as the presentations rolled on, a familiar pattern emerged: a massive list of impressive-sounding tools—Omni, Flow, Stitch, Pomelli—that exist in a state of perpetual 'coming soon'. We are seeing a shift from models that simply chat to models that act, yet the infrastructure to make these actions reliable for professional use remains shaky.

The Rise of the Agentic Workflow

The real meat of the announcements lies in Anti-Gravity 2.0. This isn't just a chatbot; it's an attempt to build a command centre for work. With features like scheduled tasks, subagents, and slash commands, Google is trying to move the interface away from the prompt box and into the workflow. The '/grill-me' command, for instance, represents a more aggressive approach to debugging than the polite clarification flows seen in Claude. It is a tool designed for people who want the machine to challenge their logic rather than just agree with it. This is the direction of travel: software that doesn't just wait for instructions but actively manages the execution of complex, multi-step projects.

The gap between a featured product and a usable one is Google's most persistent problem.

However, the creative suite—Omni for video, Flow for cinematic editing, and Stitch for UI design—reveals a different tension. Google is attempting to own the entire creative stack, from brand identity via Pomelli to the actual assembly of video assets. The ambition is to make professional-grade production accessible to anyone with a prompt. Yet, the technical reality often lags. When a tool is demonstrated in a controlled environment but fails to handle the messy, non-linear nature of real-world creative work, the hype begins to erode the trust of the very professionals it seeks to empower.

Key Takeaways for Agency Owners

Watch the shift from 'chat' to 'agentic' tools like Anti-Gravity 2.0.
Prioritise tools that offer aggressive error checking (slash commands) over polite assistants.
Be sceptical of video generation tools until they solve character consistency in Flow.
The real value is in the integration of AI into existing workspaces, not standalone apps.

Ultimately, the 2026 announcements suggest that the 'intelligence' part of AI is maturing, but the 'utility' part is still struggling. We are being promised a world where agents handle our scheduling, our coding, and our brand design, but for now, we are still largely stuck in the phase of managing the tools that are supposed to be managing us. The winners won't be the ones with the flashiest demos, but the ones who actually close the gap between the keynote stage and the user's desktop.

Key Takeaway

The era of the agent is arriving, but the reliability of these tools is still being written in the gap between demo and deployment.

02 — The Marginalian

The Biological Cost of Faith: Lessons from the Antarctic

What the survival strategies of penguins teach us about the human experience of separation.

By Maria Popova · 8 min read

Editor's note: A meditation on how endurance is not just a physical trait, but a psychological necessity.

Separation is a biological stressor. When a loved one is absent, whether through death or mere distance, the brain processes this as a fundamental disruption to safety. We often call the effort required to endure this distance 'faith'. It is the mental work of maintaining a connection to something that is no longer present. This is not a soft, poetic concept; it is a survival mechanism. For humans, it manifests in the patience of a spouse waiting for a sailor, or a parent watching a child grow into independence. It is the refusal to let the silence of absence become the reality of abandonment.

The Penguin's Stoicism

Nowhere is this more extreme than in the life cycle of the penguin. In the brutal environment of the Antarctic, survival depends on a highly coordinated, yet frequently interrupted, partnership. When a female penguin lays her egg, the male must take over incubation duties, shielding the egg with his brood patch against temperatures that would freeze it in a minute. He does this while fasting for months, standing in a huddle of other males to conserve heat, waiting for the female to return from the sea. This is a period of total vulnerability. If the male loses his resolve or his physical strength, the lineage ends.

To endure separation is to perform an act of psychological labour that we call faith.

The male penguin's devotion is not merely a romantic notion; it is a calculated biological risk. He survives on bile and stomach secretions, a physiological feat that allows him to remain a stationary guardian while his mate recovers her strength in the distant ocean. This cycle of extreme separation and eventual reunion is the engine of their species. It requires a level of trust that is almost incomprehensible to creatures that live in more stable, immediate environments. The penguin does not just endure the cold; it endures the uncertainty of the mate's return.

The Components of Endurance

Physical sacrifice: The ability to sustain oneself on minimal resources.
Psychological presence: Maintaining the 'idea' of the partner despite their absence.
Social cohesion: Using the group (the huddle) to survive individual vulnerability.

If we look at this through a wider lens, the penguin's struggle mirrors the human struggle with time and distance. We live in an age of instant connectivity, yet we still face the fundamental problem of being separate individuals. The penguin reminds us that true connection often requires the ability to withstand the void. It suggests that strength is not found in the absence of hardship, but in the capacity to remain steadfast when the thing you rely on is out of sight.

Key Takeaway

Faith is the psychological infrastructure that allows life to persist through periods of separation and uncertainty.

03 — The Marginalian

The Death of the Fact: Virginia Woolf and the Architecture of Truth

Why objective reality is a myth and how we actually construct meaning.

By Maria Popova · 10 min read

Editor's note: A challenge to the idea that truth is just a collection of data points.

There is a common mistake in how we approach knowledge: the belief that truth is something that emerges once you have gathered enough facts. We treat reality like a puzzle, assuming that if we collect enough pieces—gravity, light, historical dates—the picture of 'truth' will eventually reveal itself. But this is a fallacy. Facts are merely the static components of a world that is actually lived, felt, and perceived. Objective reality exists, but it is useless to us without the lens of human experience. Truth is not a destination reached by the intellect; it is a way of perceiving the world through our entire being.

Beyond the Intellect

Virginia Woolf argued forcefully against the idea that truth is the domain of the cold, rational mind. To her, the pursuit of truth required more than just logic; it required the 'amusements, the tendernesses, and the frivolities' of life. If we strip away our sensory experiences, our music, and our social connections in the name of pure objectivity, we don't find truth—we find a void. Woolf suggests that truth is multifaceted and disguised. It comes to us through the way we hear a piece of music or the way we feel the sun on our skin. It is a holistic phenomenon, not a mathematical one.

Truth is not the end we reach, but the manner in which we reach it.

This perspective shifts the focus from 'what' we know to 'how' we know it. If truth is a pattern hidden behind the 'cotton wool' of everyday perception, then our task is not to collect more data, but to sharpen our faculties of perception. We are not just observers of the world; we are participants in it. Woolf’s vision is one where the world is a work of art, and we are the words and the music that give it form. To seek truth is to seek a deeper alignment with the patterns that govern existence, a pattern that is felt as much as it is understood.

How to Approach Truth

Reject the idea that facts alone equal understanding.
Engage the senses and emotions as valid tools of inquiry.
Look for patterns rather than isolated data points.
Recognise that the observer is part of the observed system.

In an era of information overload, where we are drowning in 'facts' and 'data', Woolf's insight is more relevant than ever. We have more information than any generation in history, yet we seem further from the truth. This is because we have mistaken the accumulation of data for the attainment of wisdom. We are looking at the pieces of the puzzle while ignoring the pattern that connects them. Real truth requires the courage to look beyond the surface and engage with the complexity of the lived experience.

Key Takeaway

Truth is not a collection of facts, but a pattern of experience perceived through our entire being.

04 — Stratechery

The Agentic Web: Who Owns the Value of Content?

An interview with Parag Agarwal on the economic shift toward an automated internet.

By Stratechery · 15 min read

Editor's note: As agents begin to browse the web for us, the traditional economics of the internet are collapsing.

We are entering the era of the 'agentic web', a fundamental shift in how information moves across the internet. For decades, the web has been designed for human eyes: we click links, we read articles, we watch videos. But as AI agents become our primary interface for the digital world, they will be the ones consuming content, not us. This creates a massive economic problem: if an agent reads an article to answer a user's question, the user never visits the original website, and the publisher never sees the ad revenue or the subscription opportunity. The traditional value exchange of the web is breaking.

The Incentives Problem

Parag Agarwal, founder of Parallel, argues that we must find new ways to incentivise content creation in a world where humans are no longer the primary audience. If creators cannot reach humans, they cannot survive. The current model relies on attention—the ability to capture a human gaze. In an agentic world, 'attention' is replaced by 'utility'. An agent doesn't care about a catchy headline; it cares about the accuracy and structure of the data. This shift threatens to devalue everything that makes human content unique: its voice, its perspective, and its emotional resonance.

If agents consume everything, the incentive to create anything disappears.

The challenge for the next generation of platforms is to build protocols that allow agents to pay for the information they use. We need a way to track 'value' that isn't tied to a human click. This might involve micro-payments, data licensing agreements, or new types of reputation systems. Without these, the web risks becoming a desert of AI-generated sludge, where machines scrape other machines in a feedback loop that provides no new information to anyone.

Potential Solutions for the Agentic Web

Micro-payments for real-time data access by agents.
Verified human-originity tokens to distinguish content.
New subscription models that cover 'agentic access'.
Direct API-based licensing for high-quality datasets.

The transition will be painful. Many of the giants of the current web will struggle to adapt to a world where their primary asset—human attention—is being bypassed. The winners will be those who can provide high-fidelity, structured, and verifiable information that agents actually need to function. The web is moving from a place of browsing to a place of querying, and the economic rules of engagement are being rewritten in real-time.

Key Takeaway

The shift from human-centric to agent-centric browsing requires a total redesign of how we value and pay for information.

05 — Simon Willison

Datasette Agent: Bringing Intelligence to the Data Stack

How Simon Willison is turning raw databases into conversational partners.

By Simon Willison · 7 min read

Editor's note: A practical look at how LLMs are finally becoming useful for data analysis.

For years, the gap between natural language and structured data has been a significant barrier to entry for non-technical users. You can ask a chatbot a question, but getting it to write a precise SQL query that runs against a specific, local database has always been a struggle. Simon Willison's new release, Datasette Agent, aims to bridge this gap. By combining the extensibility of Datasette with the reasoning capabilities of modern LLMs, he has created a system where you can simply ask questions of your data and get immediate, actionable answers.

The Power of Extensibility

What makes Datasette Agent different from a generic AI assistant is its plugin architecture. It isn't just a wrapper around an LLM; it is a tool that can be extended to do specific things. With the addition of plugins like `datasette-agent-charts`, the agent can not only query data but also visualise it using Observable Plot. Other plugins allow for image generation or code execution in sandboxed environments. This modularity means the agent can evolve from a simple query tool into a full-scale data scientist, capable of performing complex analysis and presenting it in multiple formats.

The moment LLM and Datasette finally come together is here.

The technical foundation relies on the increasing ability of open-weight models to handle reliable tool calls and SQL generation. Willison notes that even smaller, faster models like Gemini 3.1 Flash-Lite are more than capable of handling these tasks reliably and cheaply. This is a crucial development: it means that powerful data agents don't require massive, expensive compute; they can run efficiently on local hardware or via low-cost APIs, making them accessible to individual developers and small teams.

Key Features of Datasette Agent

Conversational SQL query generation.
Automated data visualisation via plugins.
Extensible plugin architecture for custom tools.
Support for both cloud and local model execution.

The implications for personal data management are significant. Willison envisions a 'Claw'—a personal AI assistant built around the data from one's own digital life. As we accumulate more data from various sources, the ability to query that data using natural language will transform how we interact with our own histories. We are moving from a world of static files to a world of living, conversational datasets.

Key Takeaway

By combining LLMs with structured data tools, we are turning databases into interactive, conversational entities.

06 — Stratechery

The Spaghetti Problem: Why Google's AI Strategy Feels Disjointed

An analysis of the tension between DeepMind's research and Google's business needs.

By Stratechery · 11 min read

Editor's note: Is Google's AI output a sign of strength or a symptom of internal misalignment?

Google I/O 2026 was a spectacle of abundance, but for those looking closely, it felt less like a cohesive vision and more like a collection of disparate parts. This is what might be called 'I/O Spaghetti'—a mess of overlapping products, competing models, and half-baked features that seem to be thrown at the wall to see what sticks. While the sheer volume of announcements is impressive, there is a palpable lack of a single, unifying thread. It feels as though Google is trying to be everything to everyone, at the expense of a clear strategic direction.

Research vs. Revenue

At the heart of this chaos is a fundamental tension within the company: the misalignment between DeepMind's research objectives and Google's commercial imperatives. DeepMind is focused on the frontier of artificial intelligence—world models, general intelligence, and scientific breakthroughs. Google, meanwhile, is a massive advertising and cloud business that needs reliable, monetisable products. When these two forces collide, the result is often a series of 'AI-powered' features that feel tacked onto existing products rather than being built from the ground up to solve specific problems.

Google is putting AI everywhere, but it's unclear if it's putting it in the right places.

This misalignment manifests in the 'launch-to-availability' gap. We see incredible demos of video generation and agentic workflows, but when users actually try to access them, they find the tools are either restricted, buggy, or entirely unavailable. This creates a cycle of hype and disappointment. For a company of Google's scale, this isn't just a minor operational issue; it's a strategic risk. It erodes the trust of developers and enterprise customers who need stability more than they need novelty.

The Symptoms of Strategic Drift

Overlapping product lines (e.g., multiple Gemini-based tools).
A focus on 'features' rather than 'platforms'.
A disconnect between keynote demos and user experience.
Internal competition between research and product teams.

To fix this, Google needs to move beyond the 'AI-first' mantra and toward an 'AI-integrated' strategy. This means making hard choices about which capabilities to prioritise and which to leave to the research labs. It means ensuring that when a product is announced, it is actually ready for the market. Until Google can reconcile the brilliance of DeepMind with the pragmatism of its product teams, its AI strategy will continue to look like a bowl of spaghetti: plenty of substance, but impossible to navigate.

Key Takeaway

Google's AI output is currently a collection of impressive but disconnected parts, reflecting a tension between research ambition and commercial reality.