Vlad Sandu

7 months ago

Agents ship in fleets now. That’s the phase change. Orchestration becomes the deployment unit; MCP/plug-ins standardize connectors; identity + audit + revocation become the control plane. Boring ops wins because fleets widen the blast radius.

7 months ago

https://t.co/1jRA6x2ejc

557

vladsandu_ retweeted

3 months ago

On the Shoulders of Giants How Distribution Giants Could Capture Agentic AI and Commoditise the Labs Dek: The next power struggle in AI is not only about who builds the best model. It is about who owns the gateway where user intent becomes action. 1. The quiet return of the platform playbook The AI race still looks, from the outside, like a contest between frontier labs. OpenAI ships a new agent surface. Anthropic pushes Claude deeper into code and enterprise workflows. Google upgrades Gemini across Search, Android, Chrome, and cloud. Apple rebuilds Siri around personal context, app actions, and on-screen awareness. That is the headline version: a model war, with intelligence as the scoreboard. But underneath the model race, an older platform playbook is returning. The companies that own the customer interface are trying to decide where intelligence actually lives. Not in theory. In the operating system, the browser, the app store, the device, the default assistant, the developer framework, and the payment rail that decides who gets paid when an agent buys something. That is the real power struggle. The giants already own the surfaces where users express intent. Now they want those surfaces to become the places where agents act. If they succeed, frontier labs may still provide extraordinary intelligence. Their models may remain valuable, necessary, and expensive. But intelligence can become an input inside someone else’s bundle. That is the complement risk. 2. The wrong hero: raw intelligence The comfortable story says the smartest model wins. That story is not wrong. It is just incomplete. Raw intelligence matters most when intelligence is scarce. When only a few labs can produce frontier capability, the model is the power centre. But technology history rarely ends with the best component owning the market forever. The PC era did not crown the best chip alone. It crowned the operating system that became the default software gateway. The web did not crown the best page alone. It crowned the search and browser layers that routed discovery. The smartphone era did not crown the best app alone. It crowned the device, store, and OS layer that governed distribution. Agentic AI raises the stakes because the output is no longer just information. Agents do not merely answer. They retrieve, compare, book, buy, write, code, file, schedule, move, and eventually negotiate. Once AI starts acting, the strategic question changes. It is no longer only about who reasons best. It becomes about who is allowed to turn user intent into action. That is where distribution owners have an advantage. They already sit closest to intent. Apple owns the iPhone, iOS, App Store, Safari, Siri, payments, device identity, and privacy posture. Google owns Search, Android, Chrome, Gmail, Maps, YouTube, Gemini, and the Shopping Graph. Microsoft owns Windows, Edge, GitHub, Microsoft 365, Teams, Azure, and enterprise identity. Amazon owns Alexa, Prime, devices, retail, payments, logistics, ads, and merchant access. Meta owns social graphs, messaging surfaces, wearables, and consumer attention. These firms do not need to beat every frontier lab at every model task. They need to make their surface the place where AI becomes useful. Once that happens, the model underneath can be selected, swapped, blended, hidden, or priced down. The interface becomes the boss. The lab becomes a complement. 3. The strategic model: the user-intent gateway The best lens for this article is not one model. It is a stack of three. The first layer is platform economics. Platforms win when they control the interaction between two or more sides of a market. In the app era, that meant developers and users. In the agent era, it means users, agents, tools, merchants, data sources, model providers, and regulators. The second layer is boundary resources. These are the APIs, SDKs, rules, review systems, permissions, and policies that let outsiders build while the platform keeps control. In the app era, boundary resources were app-store rules, software development kits, payment policies, browser APIs, and operating-system permissions. In the agent era, they become App Intents, MCP servers, agent permissions, system actions, tool registries, memory scopes, identity grants, and audit logs. The third layer is Aginaut’s power-shift lens. The simple version is this: power follows the layer agents cannot route around. When intelligence is scarce, power sits with model builders. When intelligence becomes more available, power shifts toward distribution and orchestration. When agents act at scale, power shifts again toward governance, trust, identity, permissions, and default surfaces. This is why the user-intent gateway matters. A gateway is not just a front door. It is the layer that interprets what the user wants, chooses which agent acts, decides which tools are available, sets the permissions, captures the behavioural data, and often controls payment. That gateway can sit in several places. It can be Siri. It can be Gemini in Search. It can be Copilot in Edge. It can be Alexa. It can be Meta AI in glasses. It can be ChatGPT Atlas. It can be Claude inside Chrome. It can be a future agent store. The question is not whether agents become powerful. The question is who intermediates them. 4. Apple’s move: the App Store becomes agentic Apple’s move is the most important because Apple owns the cleanest consumer distribution stack. It controls the device, operating system, store, privacy narrative, payment rails, and user trust boundary. The reported App Store move is straightforward. Apple is exploring how to welcome AI-agent apps into the App Store while preserving its security and privacy standards. MacRumors, summarising The Information, says Apple wants to incorporate AI agents into the App Store while avoiding problems caused by rogue agents deleting content or taking unwanted actions. That matters. But the stronger receipt is Apple’s WWDC 2026 developer architecture. Apple announced that App Intents now connect apps to Siri AI capabilities such as personal context understanding, app actions, and on-screen awareness. Apple describes this as making app content and capabilities more discoverable and accessible across the system. That is the real control point. Apple is not only adding AI to apps. It is defining how apps expose themselves to system intelligence. The Apple Developer guide says App Intents is the framework for connecting apps to Apple Intelligence and Siri AI. Developers can adopt App Intents schemas so their content becomes discoverable and their capabilities become available through natural language. That sentence is a platform strategy document disguised as developer guidance. The app is no longer only something a user opens. It becomes a set of actions, entities, and permissions that Siri can understand. Apple’s Siri AI announcement pushes the same point from the user side. Siri AI can use personal context, search across apps, answer questions about screen content, go to the web for current knowledge, and take actions across apps. Apple’s Apple Intelligence page says Siri can understand personal context, take action in more apps, tap broad world knowledge, and use a dedicated Siri app. This is not just a smarter assistant. It is Apple trying to make the iPhone’s action loop run through Apple first. The user expresses intent. Siri interprets it. App Intents expose eligible actions. Apple governs the rules. The App Store remains the trust perimeter. That is agentic distribution. The model underneath can matter enormously, but it does not own the gateway. Even Apple’s model partnership with Google fits this logic. Google’s January 2026 statement says Apple Foundation Models will be based on Gemini models and cloud technology, powering future Apple Intelligence features including a more personalised Siri. That is complement risk in clean form. Google supplies capability. Apple owns the user loop. The frontier model helps Apple’s surface become more agentic. The shoulders of the giant are useful, but the giant decides where you stand. 5. Google, Microsoft, Amazon, and Meta: distribution strikes back Google is playing both sides of the board. It is a frontier lab, a search monopoly, an Android platform owner, a browser owner, a cloud platform, and a consumer assistant provider. That makes its strategy especially dangerous to pure model labs. Google’s Search announcement in May 2026 says it is bringing advanced model capabilities to Search and enabling users to use agents by asking questions. The Chrome developer team also describes Gemini in Chrome on Android as a personal browsing assistant that helps users understand web content and complete tasks with Google apps like Calendar, Keep, and Gmail. That is not a side feature. It is the browser becoming an agentic shell. Google already owns the search box. Now it wants the search box to become an action box. The same thing is happening in shopping. Amazon describes “agentic shopping” as a unified experience combining Rufus, its shopping assistant, with Alexa+, its personalised AI assistant across hundreds of millions of devices. Commerce is central to agentic AI because an assistant that can recommend, compare, and buy sits directly on the conversion layer. The old advertising funnel becomes an agent-mediated decision loop. The owner of the shopping surface gains power over what gets shown, compared, trusted, and purchased. Microsoft is attacking the same opportunity through work. Edge already positions Copilot as the browser layer that compares, summarises, remembers, and helps users finish tasks without leaving the browser. Microsoft’s Edge release notes say browsing with Copilot introduces agentic browsing for enterprises, letting Copilot navigate websites, fill information, and complete multi-step tasks for users. The bigger move is in Microsoft 365. Microsoft describes Copilot as the interface through which users interact with agents inside Microsoft 365. It says agents are like apps on an AI-powered interface, while Copilot is the interface for interacting with them. That is a remarkably explicit platform claim. Microsoft is not only selling agents. It is selling the surface that organises agents. Meta’s move is more embodied. Reuters reported on 23 June 2026 that Meta and EssilorLuxottica launched a cheaper range of AI smart glasses starting at $299, part of Meta’s investment in personal intelligence through wearable technology. Meta wants the agentic interface to sit where perception happens. Not only in a chat box. Not only in a browser. On the face. Apple owns the phone. Google owns search. Microsoft owns enterprise workflow. Amazon owns shopping. Meta wants the camera, microphone, and social graph to become its route into personal AI. Different companies. Same playbook. Own the place where the user’s intent becomes legible. Then make models compete to serve that place. 6. The frontier labs’ counter-move The frontier labs understand the danger. They are not trying to remain passive model suppliers. OpenAI’s Operator was an early signal. OpenAI described Operator as an agent that can go to the web and perform tasks by typing, clicking, and scrolling through its own browser. That is a direct move into the action layer. The model does not merely answer. It uses a browser. The Responses API and Agents SDK continue the same move. OpenAI says the Responses API supports built-in tools such as web search, file search, and computer use, designed to connect models to the real world for task completion. Then came Atlas. OpenAI describes ChatGPT Atlas as a new browser with ChatGPT built at its core. That is not an accident. A browser is a distribution surface. OpenAI cannot rely forever on being a tab inside Chrome, Safari, Edge, or mobile apps. So it is trying to own a shell. The Jony Ive move is the hardware version of the same counter-strategy. OpenAI announced that io Products merged with OpenAI in 2025, while Jony Ive and LoveFrom assumed deep design and creative responsibilities across OpenAI. That move says OpenAI knows the model alone is not enough. If Apple, Google, Microsoft, and Meta own the surface, OpenAI needs a surface of its own. Anthropic’s counter-move is less consumer-glossy but strategically serious. Anthropic introduced the Model Context Protocol as an open standard for secure two-way connections between data sources and AI-powered tools. Later, Anthropic described MCP as a universal protocol for connecting AI agents to external systems, reducing fragmentation and duplicated integration work. That is a standards play. If Anthropic cannot own every OS or device, it can try to own the way agents plug into tools. Claude Code is another surface play. Anthropic describes Claude Code as an agent that reads a codebase, edits files, and runs commands across terminal, IDE, desktop app, and browser. This is not only a coding product. It is a wedge into developer workflow. Perplexity’s Comet makes the same logic visible from another angle. Perplexity describes Comet as an AI browser that acts as a personal assistant, automating tasks, researching the web, organising email, and more. These moves all say the same thing: the frontier labs are climbing upward. They want to escape wholesale inference. They want the workflow graph, the browser, the developer shell, the device, the protocol. They want to be the place where action happens, not only the engine underneath it. 7. Historical parallels: IBM, Microsoft, Google, Apple The current AI struggle has better historical parallels than a generic “platform war.” The pattern is more specific. When a new layer threatens to become independent, the incumbent owner of the existing interface tries to bundle it, gate it, rank it, delay it, or redefine it as a feature. IBM is the earliest useful analogy. In 1969, IBM began unbundling software and services from hardware after antitrust pressure over tying software to mainframe hardware. That unbundling helped create room for an independent software market. The lesson is not simply that unbundling happens. It is that the boundary between product and complement is political, commercial, and regulatory. What counts as “part of the platform” can change. Microsoft is the cleaner browser-era analogy. The U.S. Justice Department’s findings in the Microsoft case said Microsoft had extremely large and stable share in Intel-compatible PC operating systems, protected by high barriers to entry, leaving customers without a commercially viable alternative to Windows. Microsoft’s browser fight mattered because Netscape and Java threatened to become alternative middleware layers. They could have weakened Windows as the essential gateway. Microsoft’s answer was to fold browser distribution into the operating-system advantage. That is the closest historical map for agentic AI. If an independent agent shell becomes the place where users work, search, buy, and automate, it can weaken the OS, browser, and app-store owner. So the incumbent tries to make the agent a feature. Google’s Android case adds the “open but enclosed” pattern. The European Commission fined Google €4.34 billion in 2018 for imposing illegal restrictions on Android device manufacturers and mobile operators. The Commission said Google used Android to cement the dominance of its search engine. The lesson is brutal but useful. Nominal openness does not remove platform power. Android can be open-source at one layer while Google services, search defaults, app bundles, and compatibility rules preserve control at another. That matters for agentic AI because open protocols can be enclosed by defaults. MCP can be open, yet still less powerful than whatever assistant is preinstalled. App Intents can open app actions, while Apple still governs the system invocation layer. Chrome extensions can exist, while Gemini in Chrome occupies the default surface. Apple’s App Store fights show the commerce version of the same pattern. The U.S. judge in the Epic case ruled Apple violated an order to reform App Store practices, including around external payment links and fees. The European DMA adds a broader pressure point. Apple’s own DMA page says the European Commission designated Apple as a gatekeeper for iOS, App Store, and Safari. The European Commission’s DMA developer portal says gatekeepers must effectively allow app distribution through third-party app stores or the web, and must allow developers to steer users toward alternative purchase channels. That is not old app-store trivia. Agentic commerce will reopen the same fight. When an AI agent books, buys, subscribes, negotiates, renews, or upgrades, someone will ask who gets the toll. The app-store wars become agent-store wars. 8. Strategic implications For distribution owners, the intent is clear. They want to make AI native to their existing surfaces. They want frontier labs to compete as suppliers inside those surfaces. They want developers to expose actions through approved interfaces. They want users to trust the default assistant. They want payments, identity, permissions, and memory to remain inside their governance perimeter. In short, they want to turn autonomy into a feature of the platform. Apple’s play is trust and enclosure. Google’s play is search and action. Microsoft’s play is enterprise workflow. Amazon’s play is shopping and home execution. Meta’s play is embodied attention. Each company starts from a different surface. Each wants the same thing: control over the user’s intent-to-action path. For frontier labs, the danger is obvious. The best model can still lose power if it is invoked only through someone else’s interface. A lab can be brilliant and still become a replaceable engine. That is why OpenAI’s browser, Operator, agent tooling, and hardware ambition matter. That is why Anthropic’s MCP, Claude Code, and enterprise integrations matter. They are not random product expansions. They are attempts to avoid complement status. For developers, the map gets more complicated. The old question was which platform has users. The new question is which agentic gateway can see, call, rank, approve, and monetise a capability. Developers may optimise for App Intents, Gemini actions, Copilot agents, Alexa actions, MCP servers, browser agents, or all of them. That creates a new kind of distribution tax. Not only app-store fees. Schema taxes. Permission taxes. Ranking taxes. Context taxes. Default taxes. For regulators, the old playbook will not be enough. Agentic AI makes default choice more consequential because defaults will not only answer questions. They will act. Regulators will need to ask whether users can choose default agent providers. They will need to ask whether rival agents can access system capabilities fairly. They will need to ask whether app review can suppress agentic competitors. They will need to ask whether agent memory is portable. They will need to ask whether transaction logs are auditable. They will need to ask whether first-party agents get privileged access to private APIs, payments, search data, or device context. The next antitrust fight will not only be about app distribution. It will be about action distribution. 9. What breaks next The most likely outcome is not one winner. It is layered oligopoly. Apple will remain strong wherever trust, device context, privacy, and payments matter. Google will remain strong wherever search intent, browsing, Android, maps, video, and shopping matter. Microsoft will remain strong wherever enterprise identity, documents, Teams, GitHub, and Windows workflows matter. Amazon will remain strong wherever shopping, home devices, logistics, and merchants matter. Meta will remain strong wherever social context, messaging, media, and wearables matter. Frontier labs will remain strong in reasoning-heavy, developer-heavy, enterprise-heavy, and high-trust workflows. But they will need owned surfaces or durable protocols to avoid being squeezed. The best-case scenario is open orchestration under constrained gatekeepers. In that world, users can choose default agents. Agent memory and identity become portable. MCP-like protocols remain genuinely interoperable. App stores cannot discriminate against rival agent runtimes. Browsers cannot bury alternative assistants. Agent permissions are auditable. Payment routing is contestable. The worst-case scenario is agentic enclosure. In that world, the same firms that own OSs, browsers, devices, stores, search, and social graphs also own default agents. They control tool registries, memory, ranking, payments, and what agents may do. Users get convenience. Developers get dependency. Frontier labs get margin pressure. The open web gets a new gatekeeper layer. The watchpoints are simple. Watch whether Apple turns App Intents into the default way apps become agent-readable. Watch whether Google makes Search AI Mode the default route for agentic commerce. Watch whether Microsoft turns Copilot into the enterprise agent shell. Watch whether Amazon makes Alexa+ the default shopping agent. Watch whether Meta’s glasses become a normal AI interface. Also watch the counter-moves. Watch whether OpenAI Atlas gains enough distribution to matter. Watch whether Anthropic’s MCP remains neutral or gets enclosed by dominant platforms. Watch whether regulators treat agent defaults like browser defaults. Watch whether agent payment flows trigger the next Epic-style fight. Watch whether developers optimise for open protocols or for the largest closed gateways. That will tell us whether frontier labs remain platforms, or become engines under someone else’s hood. 10. The line to remember The phrase “on the shoulders of giants” usually sounds generous. It suggests progress built on what came before. But in platform markets, the shoulders of giants are not neutral ground. They are surfaces with rules. They are toll roads. They are app stores. They are browsers. They are operating systems. They are search boxes. They are devices. They are defaults. The giants are not merely lending their shoulders to AI. They are deciding who gets to stand there, what they can see, what they can touch, and who gets paid when they move. That is the real agentic AI power play. AI was first a model war. Agentic AI is becoming a gateway war. And the gateway is where user intent becomes action.

139K

vladsandu_ retweeted

4 months ago

https://t.co/VYzy1uzb72

390K

vladsandu_ retweeted

5 months ago

https://t.co/jASbCevkg0

484K

vladsandu_ retweeted

Peter H. Diamandis, MD

@PeterDiamandis

5 months ago

https://t.co/QMLUvLre9k

298

178

24K

vladsandu_ retweeted

6 months ago

https://t.co/i7TEf63SZp

vladsandu_ retweeted

8 months ago

https://t.co/cLDWG8Ju47

579

vladsandu_ retweeted

7 months ago

https://t.co/1jRA6x2ejc

557

vladsandu_ retweeted

7 months ago

https://t.co/3t9Crw14Co

971

vladsandu_ retweeted

7 months ago

https://t.co/AEndAOqcx5

171K

vladsandu_ retweeted

6 months ago

https://t.co/rs0fQ8eSV9

58K

vladsandu_ retweeted

6 months ago

https://t.co/NVrCt1xYEp

149K

vladsandu_ retweeted

7 months ago

https://t.co/30Jh6O4qEz

51K

7 months ago

Default-deny is starting to look like a growth strategy. As per AI BEACON #05 - the quiet shift is that incentives are moving inside the answer path. Once monetisation is in-band, “policy�� becomes routing rules.

7 months ago

https://t.co/3t9Crw14Co

971

8 months ago

The interface won. The model just works there.

8 months ago

Agents don’t scale by IQ. They scale by where they’re allowed to click.

8 months ago

https://t.co/cLDWG8Ju47

579

8 months ago

We are speedrunning from “nice answer” to “Stripe + IAM + Jira inside a chat bubble.” MCP Apps = buttons/forms in-chat, Copilot/Claude = agents where work happens, Mastercard = agent-native checkout… and users are basically saying: cool, but give me an undo button and liability clarity.

8 months ago

https://t.co/eHunFeuzc1

826

101

vladsandu_ retweeted

8 months ago

@NathanLands

8 months ago

“Alignment” is starting to look a lot like IAM in a hoodie. Once agents can click purchase or touch a ledger, the constraint isn’t IQ . It’s authority + rollback. MCP + audit + quarantine lanes are the real story this week.

8 months ago

https://t.co/r6sOj9WJNA

vladsandu_ retweeted