Agentic web infrastructure is the open protocol layer that enables autonomous AI agents to discover, verify, and coordinate with each other across organizational boundaries - replacing the proprietary walled gardens that currently restrict enterprise AI interoperability. According to Gartner, by 2028 over 33% of enterprise software will include agentic AI, up from less than 1% in 2024, making open infrastructure a strategic imperative.
The shift from conversational chatbots to autonomous agents marks the beginning of a fundamental reorganization of the digital economy. As organizations move beyond simple prompts, they are discovering that the current agentic web infrastructure is surprisingly fragile, governed largely by proprietary platforms that restrict interoperability. We are entering an era where the internet will host not millions of documents, but trillions of autonomous agents. These agents will negotiate, delegate, and migrate across hosts in milliseconds - a reality that requires a new layer of open infrastructure for discovery, identity, and coordination.
The AOL era of agentic web infrastructure
To understand the current state of AI agents, we must look back at the late 1990s. During the AOL era, users lived within a gated directory. You installed software from a CD, and your entire digital experience was curated and controlled by a single company. It was a safe, functional, but ultimately limited "walled garden." What followed was the open web - a permissionless environment where any browser could talk to any website via open protocols like HTTP and HTML.
Today, AI agents are in their "AOL phase." Most agents are built and confined within proprietary stores or closed ecosystems. They can only interact with other tools and agents if they share the same vendor. This creates a massive bottleneck for scaling AI agent operations. If an agent at Company A cannot discover or transact with an agent at Company B without a pre-existing enterprise integration project, the true potential of an autonomous economy remains locked. The transition from these closed gardens to an open agentic web is the most critical architectural shift facing operations leaders today.
The architecture of discovery: beyond the Domain Name System
On the human web, we use the Domain Name System (DNS) to map human-readable names to IP addresses. However, DNS was designed for static documents, not autonomous actors. In an agentic economy, an address is not enough. An agent needs to know what another agent is capable of, what tools it can access, what security rules it follows, and how to initiate a handshake.
Research into decentralized architectures highlights the necessity of a dedicated discovery layer, such as the Nanda Index. This acts as a shared registry where agents publish an "agent card." This card contains the machine-readable metadata required for one agent to understand the "functions" of another.
When an agent seeks a collaborator - for example, a logistics agent looking for a local delivery agent - it queries the index. Instead of a simple IP address, it receives an updated set of agent facts. This resolution is adaptive; the information returned can change based on who is asking and what level of access they are permitted. This ensures that while agents are discoverable, their internal logic and private data remain shielded behind a gateway. Organizations building multi-agent AI architectures must account for this discovery layer from day one.
Verifiable trust through the Agent Facts Record
In a world where agents operate on behalf of corporations, trust cannot be assumed; it must be verified. The "Agent Facts Record" is the cornerstone of this trust layer. It is a signed, immutable record that provides several critical pieces of information:
- Provenance: Who built the agent and which organization owns it?
- Capabilities: What tasks is the agent authorized to perform?
- Permissions: What data or systems is the agent allowed to touch?
- Reachability: Where should messages be sent to initiate coordination?
Before any autonomous connection is made, agents can perform a "fact check" on each other. This prevents a common failure mode in current AI experiments: the "Shadow AI" problem, where ungoverned agents interact with sensitive systems without a clear audit trail. By moving toward a standardized facts record, organizations can ensure that every autonomous action is backed by a verifiable identity that passes enterprise procurement and security standards.
Sovereign agent runtimes: the foundation of agentic web infrastructure
If agents are to do real work, they need access to real tools: CRM systems, financial ledgers, and private communication channels. This raises a critical question of control. Entrusting these capabilities to a third-party SaaS platform creates significant security and consistency risks. This is why the industry is shifting toward sovereign, self-hosted runtimes.
A managed instance approach allows an organization to run its agents on its own infrastructure - whether that is a private cloud or a secure VPC - ensuring that the "agent loop" stays within the company's governance boundary. When an agent is self-hosted, the organization owns the memory, the logs, and the integration secrets.
This sovereignty is not just about security; it is about operability. A production-grade agent system needs to be persistent, scheduled, and auditable. It cannot be an ephemeral script running on a developer's laptop. It must be part of the company's core infrastructure, as reliable and governed as its primary database. See how Ability.ai builds sovereign agent infrastructure that keeps your agents under your control.
Per-agent economics and the sleep-and-wake architecture
One of the primary barriers to the "trillions of agents" vision is cost. In traditional cloud computing, keeping thousands of agents "alive" and waiting for tasks is prohibitively expensive. This is why the development of per-agent economics is so vital.
New hosting models utilize a "sleep-and-wake" architecture. In this model, an idle agent does not consume active compute resources. It sits in a dormant state until a message arrives in its "message box." The infrastructure then wakes the agent, it processes the task, and returns to sleep.
This shift moves the cost conversation away from "per-seat" SaaS subscriptions toward "per-agent" utility pricing. For a scaling company, this is the difference between a project that is a cost center and an autonomous system that scales horizontally without a linear increase in headcount or software fees. It allows a business to deploy a dedicated agent for every single customer or every individual workflow without breaking the cloud budget. Teams evaluating AI vendor lock-in risks should consider per-agent economics as a key criterion.
Orchestrating the bazaar: stress-testing agentic web infrastructure at scale
How do we know if these decentralized protocols will actually hold up under the pressure of a real-world economy? The answer lies in massive-scale simulation. Project Nanda's "Nanda Town" provides a sandbox for testing how thousands of agents interact in what is being called the "Bazaar Era."
In this simulated environment, researchers can observe agents as they:
- Negotiate prices: Buyers and sellers use autonomous reasoning to find equilibrium in real-time.
- Submit ballots: Testing decentralized voting and consensus mechanisms for corporate governance.
- Coordinate supply chains: Managing handoffs between multiple agents across different "organizations" to ensure a task is completed without human intervention.
The research breaks the agentic web into 12 distinct layers - including transport, identity, auth, trust, payments, and memory. By isolating these layers, we can stress-test specific protocols. For example, we can see how an agent recovers when its primary payment gateway fails or how it handles "poisoned" messages from a malicious actor. This "simulation-first" approach is becoming a requirement for enterprise AI agent governance, allowing companies to prove the value of a system before it becomes load-bearing for the business.
Strategic implications for operations leaders
The move toward an open agentic web is not a technical trend; it is a governance evolution. For CEOs and COOs, the primary risk is not "missing out on AI," but becoming trapped in a proprietary ecosystem that does not talk to the rest of the world.
The path forward requires a solution-first approach. Rather than building fragmented experiments that live in isolated silos, organizations should look toward building sovereign agent systems on open infrastructure that allows for future interoperability - ensuring that as the "bazaar" grows, your organization's agents are ready to participate, negotiate, and thrive.
By focusing on the infrastructure of discovery and trust today, companies can move from the "AOL era" of AI into a future where autonomous agents are a reliable, governed, and highly profitable extension of the workforce. The transition will be led by those who prioritize data sovereignty and per-agent operability, turning the "Shadow AI" of today into the sovereign infrastructure of tomorrow.
If your team is exploring agentic web infrastructure for production deployment, explore how Ability.ai delivers governed AI solutions without platform lock-in or subscription fees.
