The Agentic Web: Disassembling Software for a New Era

TL;DR: AI is disrupting traditional monolithic applications by enabling AI agents to dynamically assemble data and services to achieve specific goals. The result, a collection of data sources and best-of-breed microservices (discrete apps) operated by AI agents. Two new technologies, MCP and https://www.a2aprotocol.org/en, make this vision a reality.

In my earlier post, AI is Eating Software, I argued that AI is dismantling traditional software by shifting focus from rigid, UI-driven workflows to the flexible, powerful engines beneath. This vision is now accelerating with the rise of the Agentic Web—a headless ecosystem of disassembled engines, data, metadata, and AI agents that interact dynamically to solve complex problems. Two new protocols, MCP servers (from Anthropic) and agent-to-agent (A2A) communication (from Google) are supercharging this trend, enabling AI to act as a full-fledged client and unlocking the long-tail of use cases that monolithic apps could never address. In this post, I’ll explore how large language models (LLMs), fine-tuned models, and these cutting-edge technologies are reshaping the software industry by dismantling traditional architectures and fostering a microservices-driven future.

AI as a Full-Fledged Client

The core idea from AI is Eating Software is that AI agents bypass the constraints of traditional user interfaces, interacting directly with software engines via APIs or database queries. Unlike fixed UI workflows, which limit users to predefined paths, these engines offer immense flexibility. For example, a CRM’s UI might streamline logging a sales call, but an AI agent can query the underlying database to analyze patterns, automate follow-ups, or create custom reports—unlocking niche use cases that serve the long-tail of business needs. Technologies like MCP and A2A amplify this capability by providing universal access to the ecosystem, of applications and databases. This enables agents to operate and collaborate at scale, accelerating the shift away from UI-centric software.

LLMs and Fine-Tuned Models: The Backbone of the Agentic Web

Large language models (LLMs), trained on vast public datasets, provide the general-purpose reasoning and task-execution capabilities that power the Agentic Web. These models are the foundation for agentic interactions, but their true potential is realized when fine-tuned with proprietary data tailored to enterprise needs. Here’s how:

• Domain Expertise and Use-Case Specific Nuances: Fine-tuned models leverage internal data from databases and SaaS applications via Retrieval-Augmented Generation (RAG). This enables agents to master industry-specific contexts—like regulatory compliance in healthcare or inventory optimization in retail—and address use-case specific nuances with precision.
• Access to Proprietary Data: By securely integrating enterprise data, fine-tuned models deliver personalized, real-time solutions grounded in proprietary insights.
• Synthetic Testing: Fine-tuning supports the generation of synthetic data, accelerating model evolution and evaluation without compromising sensitive information.

MCP servers (e.g., Anthropic’s infrastructure) are critical enablers here. These two-way services allow agents to connect with SaaS apps and databases, gathering data, metadata, and insights while operating their underlying engines. Constrained by Role-Based Access Control (RBAC), MCP ensures secure, governed interactions. For example, an agent might use MCP to pull customer data from a SaaS platform, analyze it, and trigger automated actions—all while adhering to access permissions. This seamless integration accelerates the trend of AI eating software by enabling agents to directly manipulate engines, bypassing traditional UI bottlenecks.

Agentic Operations: Iterating Toward Optimal Outcomes

AI agents in the Agentic Web are not passive data collectors—they act and optimize. Operating within RBAC constraints, agents gather data, metadata, and insights, then take iterative actions to achieve user-defined goals. For instance, an agent tasked with boosting e-commerce conversions might query a database for user behavior, test personalized offers, and refine its approach based on engagement metrics. This iterative feedback loop, powered by MCP’s ability to connect agents with engines, ensures continuous improvement and drives efficiency at scale.

Agent-to-Agent (A2A) Communication: The Power of Collaboration

The Agentic Web thrives on collaboration, and A2A communication (e.g., Google’s framework) is a game-changer. Imagine a line of ants working together to carry a large piece of food: each ant handles a specialized task while coordinating with others. Similarly, A2A enables agents to communicate, negotiate, and divide tasks autonomously. For example, one agent might extract financial data from an ERP system, while another optimizes budget allocations based on that data. This distributed problem-solving, accelerated by A2A, amplifies the vision of AI eating software by enabling scalable, collaborative workflows that reduce human intervention and enhance efficiency.

Disrupting the Software Industry

As I outlined in AI is Eating Software, AI’s ability to interact directly with engines is dismantling traditional software paradigms. MCP and A2A technologies are accelerating this disruption in three key ways:

• Minimized Investment in UI/UX: With agents accessing engines via APIs, the need for polished, monolithic user interfaces shrinks. Developers can prioritize building robust, flexible engines, reducing the costly focus on UI/UX design.
• Disassembly of Monolithic Apps: Traditional apps are fragmenting into microservices, each addressing specific functions. Agents, using MCP to connect with these services, can leverage the strengths of multiple platforms—combining, say, Salesforce’s CRM with HubSpot’s marketing tools—to create tailored workflows without platform lock-in.
• New Opportunities for Innovation: Developers no longer need to build full applications to compete. With A2A enabling seamless integration, they can create specialized engines to fill gaps in larger platforms. For example, a startup could build an engine to enhance a specific feature of a massive SaaS app, selling directly into its user base. This is like a million ants eating a lunch, with each ant tackling a small, specialized piece—made possible by A2A’s collaborative power.

An Example of The Disruptive Power

Imagine you have a brilliant idea for a new marketplace called "DealMatch" that connects buyers and sellers of second-hand luxury goods (e.g., designer handbags, watches, and jewelry) using a cool new AI engine. You realize that success is about 1% tech and 99% building critical mass of buyers and sellers. So, instead of building an entire marketplace to use your engine, you just expose it to agents using the huge marketplaces like eBay, Poshmark or The RealReal. Now you have a huge installed base to tap into instantly. You expose it to agents via MCP servers for the marketplaces and A2A. The agents recognize your value and start using it. Your market share and revenues spike, without you investing years and millions of dollars trying to build a marketplace and establish a userbase of buyers and sellers. 

The Future of the Agentic Web

The Agentic Web, supercharged by MCP and A2A, is transforming the software landscape. Businesses gain customizable, cost-effective solutions by adopting only the microservices they need. Developers face lower barriers to entry, competing through targeted engines rather than full apps. However, challenges remain:

• Security and Governance: RBAC and secure data handling are critical as agents access sensitive systems via MCP.
• Interoperability: Standardized protocols for A2A communication are essential for seamless collaboration.
• Managing Complexity: A fragmented microservices ecosystem requires robust orchestration to avoid chaos.

Conclusion

The Agentic Web builds on the vision I laid out in AI is Eating Software, taking it to new heights with technologies like MCP servers and A2A communication. By leveraging LLMs, fine-tuned models, and these innovative frameworks, AI agents are dismantling monolithic applications and enabling a future of flexible, engine-driven microservices. This shift empowers businesses to unlock the long-tail of use cases, developers to innovate with targeted solutions, and the industry to rethink software design. The ants are marching faster than ever—embrace the Agentic Web to stay ahead in this transformative era.