The Agentic Web

How Universal Protocols Will Restructure Digital Power

## I. The Protocol Moment

We stand at an inflection point comparable to the standardization of TCP/IP or HTTP—but compressed into months rather than decades. Google's Agent2Agent (A2A) protocol, Microsoft's embrace of Anthropic's Model Context Protocol (MCP), and the rapid alignment of enterprise players signal more than interoperability improvements. They represent the emergence of a **universal substrate for autonomous coordination**: agents discovering each other through standardized cards, negotiating tasks via JSON-RPC, and executing complex workflows across organizational boundaries without human translation layers.

The technical architecture is elegant in its simplicity. MCP standardizes the "last mile"—how agents access tools, databases, and APIs. A2A standardizes the "lateral mile"—how agents communicate peer-to-peer, delegate subtasks, and coordinate multi-step processes. Together, they create what Kevin Scott calls the "agentic web": an internet where the primary traffic is machine-to-machine negotiation rather than human browsing. Crucially, agents operate with authenticated identity (via OIDC/JWT), fine-grained permissions, and observable state management—the infrastructure for autonomous economic actors, not just smart tools.

This isn't vaporware. Over 50 technology partners—Salesforce, SAP, MongoDB, Atlassian—have committed to A2A. Microsoft has embedded MCP into Windows 11, making every desktop application potentially agent-accessible. OpenAI is integrating MCP into ChatGPT. The protocol wars are over before they began; the industry has chosen **structured interoperability over proprietary moats**. The question is no longer *whether* the agentic web arrives, but *what gets destroyed* in its wake.

## II. Three Structural Disruptions

### Hypothesis 1: The Elimination of Interface Friction as Competitive Advantage

The most profound impact is the least discussed. Today's platform power rests partly on **cognitive switching costs**—the mental overhead of learning new interfaces, migrating data, reconfiguring workflows. Users tolerate Salesforce's mediocrity or Microsoft's bloat because retraining on alternatives imposes coordination costs across teams, muscle memory, and institutional processes.

Agent protocols **obliterate this moat entirely**. When your procurement agent can negotiate with any supplier's agent via A2A, your loyalty to SAP Ariba becomes irrelevant. The interface you never see can't create lock-in. Your agent doesn't care whether the backend is Oracle or a startup—it parses the same JSON, follows the same authentication flows, receives the same structured responses.

Research confirms the mechanism: MCP reduces the M×N integration problem to M+N by creating universal adapters. Any MCP client can access any MCP server. What previously required custom connectors for every platform pairing now requires a single protocol implementation. The proprietary "context flywheel"—where accumulated interaction data creates switching costs—gets replaced by portable memory standards. Your agent's learned preferences about your vendor selection criteria travel with you, not with the platform.

The implication: **User experience commodifies**. Differentiation must shift to quality of underlying service, not interface stickiness. Platforms built primarily on workflow lock-in face compression toward marginal cost pricing.

### Hypothesis 2: The Collapse of Platform Power Through Context Portability

The trillion-dollar platforms—Google, Microsoft, Salesforce—derive power from proprietary context accumulation. Your Gmail search history trains better email prioritization. Your GitHub commits train better code completion. Your CRM interaction logs train better lead scoring. This creates vertical integration: the platform that captures context owns the model that exploits it, which captures more context in a compounding loop.

Open protocols **break the vertical integration**. MCP explicitly separates context provision from model execution. Any AI application can request structured data from Google Drive, Slack, or Notion—if the user authorizes it. The context layer becomes **infrastructure, not competitive advantage**.

McKinsey's research on agentic AI architecture emphasizes "vendor neutrality" as a core design principle, with MCP and A2A explicitly preferred over proprietary protocols. The strategic imperative is avoiding lock-in as technology rapidly evolves. But vendor neutrality is a euphemism: it means **commoditization of platform infrastructure**. When enterprises can swap Salesforce for a competitor by updating agent configurations rather than ripping out integrations, pricing power evaporates.

The research literature on AI lock-in reveals the pattern clearly: organizations increasingly demand "mix-and-match" architectures to mitigate switching costs. The same forces that drove cloud standardization (avoiding vendor capture, preserving optionality) now drive protocol adoption. But cloud commoditization took 15 years; protocol commoditization happens at software speed.

Consider the attention lemons problem formalized in recent economic modeling: users' rational decisions to delegate browsing to agents create negative externalities for publishers, as synthetic agent traffic dilutes attention quality sold to advertisers. Publishers face fragmentation—block agents and lose traffic, allow agents and lose ad revenue. The only sustainable equilibrium involves **agents as the primary interface**, which collapses platform power built on attention intermediation.

### Hypothesis 3: The Obsolescence of Attention-Based Advertising

Digital advertising—a $360B market—rests on a simple mechanism: capturing human attention, exploiting cognitive biases (urgency, social proof, authority), and converting eyeballs to clicks. Banner ads, sponsored search results, native content—all optimized for human psychology, not rational evaluation.

Agent-mediated commerce **breaks the entire value chain**. When your shopping agent compares hotel options, it doesn't "see" banner ads. It parses structured data: pricing tables, availability calendars, review distributions, cancellation policies. Empirical research on agent decision-making confirms this: GPT-4 and Claude respond to text-based, keyword-rich structured content while systematically ignoring or undervaluing visual advertising. Machines engage with what they can parse—semantic markup, schema.org annotations, API-provided metadata.

The Perplexity AI vision makes this concrete: merchants compete for agent attention *in the backend*, bidding for inclusion in shortlists the user never sees. "The user never sees an ad. Merchants are not competing for users' attention; they're competing for the agents' attention." This isn't speculative—Perplexity's ARR grew from $3M to $70M in 2024, demonstrating market validation of AI-intermediated discovery.

What replaces advertising? **Structured optimization**. Instead of A/B testing ad creative for emotional resonance, brands optimize machine-readable content: product specifications, pricing structures, API response times, schema completeness. Instead of brand awareness campaigns, investments in trusted data sources that agents consult. Instead of persuasive copywriting, semantic clarity and contextual alignment.

The shift resembles SEO's evolution—but faster and more total. Early SEO meant gaming keyword density; mature SEO meant quality content. Agent optimization means becoming a **trusted information source** that LLMs cite during reasoning. The business model transitions from attention capture to authority establishment. Gartner predicts a 25% drop in traditional search volume as AI search ascends—but the reality is more radical: search doesn't decline, it becomes **invisible**, handled by agents in background processes.

The advertising industry recognizes the threat. WPP's CEO admits AI "totally disrupts our business." But the response—using AI to generate ad creative faster—misses the point. The problem isn't production efficiency; it's that **the target audience disappears**. You can't advertise to an agent optimizing on price and delivery time. You can only *be* the best option by those metrics.

## III. Strategic Implications

### For Incumbents: Defend Defensible Ground

The platforms built on interface lock-in—Salesforce, Workday, ServiceNow—face an existential challenge. Their moats are cognitive switching costs and integration complexity. Protocols eliminate both simultaneously. The strategic response cannot be "improve the UI" or "deeper integration"—those strategies assume human users remain the primary interface.

**Three defensive plays have merit:**

1. **Own the agent orchestration layer.** Microsoft's strategy is instructive: embrace open protocols (MCP, A2A) while controlling the orchestration platform (Azure AI Foundry, Copilot Studio). Commoditize context access, monetize workflow complexity. The value shifts from "we have your data" to "we coordinate your agents."

2. **Become the trusted data provider.** If advertising becomes authority-based, invest heavily in becoming *the* canonical source for your domain. Bloomberg's strategy translates: own proprietary datasets, maintain gold-standard accuracy, charge for API access that agents rely on. You can't stop agents from querying competitors, but you can become the source they weight most heavily.

3. **Migrate revenue models before they collapse.** Don't wait for attention-based advertising to slowly decline. Proactively shift to transaction fees, SaaS subscriptions, or data licensing. The transition window is narrow—perhaps 18-36 months before agent adoption reaches critical mass in key verticals (travel, e-commerce, B2B procurement).

**What won't work:** Proprietary protocols, walled gardens, or agent-blocking. The industry has collectively chosen interoperability. Swimming against that tide ensures marginalization. The cautionary tale is Internet Explorer—Microsoft's attempt to control web standards through proprietary extensions led to irrelevance once the industry standardized on Chromium.

### For Startups: Attack Newly Vulnerable Ground

The agentic web creates arbitrage opportunities wherever incumbents remain dependent on interface lock-in or attention capture.

**Highest-leverage attack vectors:**

1. **Agent-native vertical SaaS.** Build the travel booking platform, legal research tool, or procurement system designed for agent-to-agent communication from day one. No human UI except admin dashboards. Optimize for API response times, structured data quality, and protocol compliance. Incumbents burdened with legacy UI codebases and human-centric workflows can't pivot fast enough. Your advantage: **greenfield architecture** optimized for machine clients.

2. **Context portability infrastructure.** Build the tools that help enterprises extract proprietary context from incumbents and feed it to their own agent orchestration layers. Think "Plaid for agent memory" or "data liberation as a service." Incumbents profit from context lock-in; you profit from breaking it. As MCP adoption grows, enterprises will pay handsomely for tools that operationalize data portability rights.

3. **Agent-to-agent marketplaces.** Create the "app store" where agents discover each other—curated registries, reputation systems, pricing transparency. Google and Microsoft will build generic registries, but specialized verticals need tailored discovery. A marketplace for compliance agents serving financial services, or supply chain agents with verified provenance tracking, creates network effects and liquidity that generalists can't match.

4. **Semantic optimization platforms.** If advertising becomes structured content optimization, build the "SEMrush for agents"—tools that audit how well your product data ranks in agent evaluations, identify gaps in schema markup, benchmark API performance against competitors, and A/B test structured content variations. Every e-commerce brand, SaaS platform, and professional services firm will need this, and incumbents (built for human-centric marketing) lack the technical DNA.

**The timing window:** First-mover advantage in agent-native categories compounds rapidly. Network effects from agent adoption create winner-take-most dynamics in orchestration layers and marketplaces. But the window is narrow—major players are moving fast. Launch MVPs within 6 months, achieve product-market fit within 12, or risk permanent marginalization as standards congeal around early movers.

### The Meta-Strategic Insight

The agentic web isn't the next platform—it's the **unbundling of platforms**. HTTP unbundled mainframe applications into networked services. Mobile unbundled desktop software into app ecosystems. Protocols unbundle platforms into composable, agent-orchestrated workflows.

The winners won't be those with the best user interfaces or strongest data moats. They'll be those who recognized earliest that **the unit of competition shifts from applications to capabilities**—discrete, agent-accessible services that compose into emergent workflows. Think microservices architecture, but for entire business models.

The cognitive shift required: stop optimizing for human attention and start optimizing for machine legibility. Stop building moats from switching costs and start building moats from **capability excellence**—being genuinely the best at your atomic function, not just the stickiest. The agentic web rewards the specialized and composable, punishes the bundled and proprietary.

We've seen this movie before—with the web, with mobile, with cloud. Those who adapt to the new substrate thrive. Those who defend the old paradigm become footnotes. The difference this time: the cycle completes in quarters, not decades.

---

*The agentic web is not a distant vision. The protocols exist. The industry has aligned. The agents are already negotiating. The question is whether you're building for the world that's arriving, or defending the world that's disappearing.*