Your company’s AI agent can schedule meetings, draft proposals, and query your CRM. What it cannot do is negotiate contract terms with your supplier’s agent, coordinate logistics with your shipping partner’s agent, or verify credentials with your customer’s compliance agent. Every enterprise is building agents that work brilliantly in isolation and accomplish nothing across organizational boundaries. That is the problem the agent web solves.
The concept is simple and the analogy is precise: the world wide web connected documents across servers. The agent web connects AI agents across companies. Google’s A2A protocol provides the communication layer. Anthropic’s MCP provides the tool access layer. And a discovery mechanism called Agent Cards, hosted at /.well-known/agent.json, lets agents find and evaluate each other automatically. Fifty-plus enterprise vendors, including SAP, Salesforce, ServiceNow, and Workday, have already signed on.
Why Internal Agents Hit a Wall
The average enterprise now operates 12 AI agents, according to MuleSoft’s 2026 Connectivity Benchmark. But 50% of those agents run in silos, disconnected from the systems they need. The problem compounds when you look beyond company walls.
Consider a real procurement workflow. A purchasing agent at a manufacturer needs to source components. Today it queries the internal ERP, checks inventory, maybe drafts a purchase order. But the actual procurement process involves supplier agents that can quote prices, logistics agents that can confirm delivery windows, customs agents that can verify compliance, and payment agents that can process transactions. These agents exist at different companies, built on different frameworks, running behind different firewalls.
Without a standard way for these agents to discover each other, verify identity, negotiate terms, and exchange structured data, every cross-company integration requires custom point-to-point wiring. That is exactly where enterprise software was in the 1990s before HTTP and HTML standardized web communication.
The Numbers Behind the Gap
Gartner predicts that 33% of enterprise software will include agentic AI by 2028, up from less than 1% in 2024. Capgemini’s research shows 82% of organizations plan to integrate AI agents within three years. McKinsey estimates agentic AI could unlock $2.6 to $4.4 trillion in annual productivity gains. But most of that value requires agents that operate across organizational boundaries, not just within them.
The gap is clear: enterprises are building agents fast, but building them as islands.
The Agent Web Stack: How It Actually Works
The agent web is not a single protocol. It is a stack, and each layer solves a different problem.
Layer 1: Tool Access (MCP)
Anthropic’s Model Context Protocol, now governed by the Linux Foundation’s Agentic AI Foundation, standardizes how agents access tools and data. Think of it as USB-C for AI: one connector for any tool, any data source, any service. An agent using MCP can read from GitHub, query a database, and call a REST API through a single standardized interface.
MCP operates within organizational boundaries. It answers the question: “How does my agent connect to my tools?”
Layer 2: Agent-to-Agent Communication (A2A)
Google’s Agent-to-Agent protocol, launched in April 2025 with 50+ partners, standardizes how agents from different vendors talk to each other. A2A answers a different question: “How does my agent collaborate with your agent?”
The protocol supports task delegation (send a task to a remote agent), streaming responses (real-time status updates via Server-Sent Events), multi-turn conversations (back-and-forth negotiation between agents), and multimodal content exchange (text, files, structured data).
A2A’s launch partners read like an enterprise vendor directory: Atlassian, Box, Cohere, Intuit, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG, and Workday.
Layer 3: Discovery (Agent Cards)
This is the layer that makes the web analogy literal. Just as websites publish robots.txt to tell crawlers what they can access, agents publish Agent Cards at /.well-known/agent.json to advertise their capabilities. An Agent Card is a JSON file that describes:
- What the agent can do (capabilities)
- How to authenticate (OAuth 2.0, API keys)
- What input/output formats it supports
- Where to send requests (endpoint URLs)
An orchestrator agent can crawl these cards to discover available agents, evaluate their capabilities, and decide which ones to engage for a given task. No hardcoded integrations needed.
What Cross-Company Agent Networks Look Like in Practice
This is not theoretical. Several enterprise platforms are already building cross-organizational agent workflows.
Salesforce Agentforce: Customer Service Across Company Lines
Salesforce’s Agentforce platform enables agents that span organizational boundaries. A customer service agent handling a warranty claim can autonomously escalate to a supplier’s return-processing agent, coordinate a replacement through a logistics partner’s shipping agent, and update the customer record, all without a human touching the ticket. The agents authenticate via A2A, exchange structured task objects, and track completion status across company lines.
SAP Business Network: Procurement Goes Autonomous
SAP’s business network connects 5.5 million companies. With agent capabilities built into SAP Joule, a buyer’s procurement agent can discover supplier agents on the network, request quotes, compare terms, negotiate pricing, and place orders. The human reviews exceptions. The agents handle the routine.
ServiceNow: Cross-Platform Incident Response
When a cloud outage hits multiple customers, ServiceNow agents can coordinate with infrastructure provider agents to diagnose root causes, estimate resolution times, and communicate status updates. Instead of humans in three companies trading emails and Slack messages, agents exchange structured incident data in real time.
The Unsolved Problems: Trust, Liability, and Regulation
Building the agent web’s technical infrastructure is the easy part. The hard problems are all human.
Identity and Trust
When Company A’s agent receives a task request from Company B’s agent, how does it verify that the request is legitimate? OAuth 2.0 and OpenID Connect are being adapted for agent-to-agent authentication, but the trust model is fundamentally different from human SSO. A human logs in once and works within a session. Agents chain requests across multiple systems, and each system in the chain needs to verify the original authorization context, not just the immediate caller’s service account.
The web solved a similar problem with SSL/TLS certificates and certificate authorities. The agent web will need analogous trust infrastructure: agent identity certificates, capability attestation services, and audit trail standards.
Liability
If your procurement agent negotiates a bad deal with a supplier’s agent, who is responsible? The buyer? The supplier? The agent vendor? Current contract law assumes humans make decisions. When two autonomous agents close a deal at 3 AM, the legal framework is genuinely unclear. The EU AI Act addresses AI system transparency and risk classification, but it does not yet cover agent-to-agent transactions across organizational boundaries.
Data Privacy
GDPR requires a legal basis for data processing. When your agent shares customer data with a partner’s agent to fulfill a request, both sides need clear data processing agreements. The agent web creates data flows that are faster, more numerous, and harder to audit than human-mediated processes. Privacy by design needs to be built into the protocol layer, not bolted on after deployment.
The Web Analogy Is Not Just Marketing
The parallels between the early web and the emerging agent web are structural, not cosmetic.
| World Wide Web | Agent Web |
|---|---|
| HTML (content format) | Agent Cards (capability description) |
| HTTP (transport protocol) | A2A (agent communication protocol) |
| URLs (addressing) | Agent endpoints + /.well-known/agent.json |
| DNS (discovery) | Agent registries and catalogs |
| robots.txt (permissions) | Agent Cards (auth and capabilities) |
| SSL/TLS (trust) | OAuth 2.0 + agent identity certificates |
| Search engines | Agent marketplaces and directories |
| Hyperlinks (connections) | Agent-to-agent task delegation |
The early web faced the same challenges: competing standards (remember SGML vs HTML?), trust issues (commerce was “impossible” online until SSL), and governance questions (who controls the web?). The agent web will follow a similar maturation curve, likely compressed into years instead of decades.
The historical precedent is worth noting. The Foundation for Intelligent Physical Agents (FIPA) attempted to standardize agent communication in the early 2000s through IEEE. It failed because the technology was not ready and adoption was too low. The difference now is that enterprises already run agents in production, the major vendors have aligned on open protocols, and the economic incentive (trillions in productivity gains) is large enough to force interoperability.
What to Do About It Now
You do not need to wait for the agent web to fully mature. Three things you can do today:
Adopt MCP for internal tool access. If your agents connect to internal tools through custom integrations, start migrating to MCP. When the agent web arrives, your agents will already speak the right language for tool access.
Evaluate A2A for partner integrations. If you have B2B workflows that involve structured data exchange with partners (procurement, logistics, compliance verification), pilot A2A for one workflow. Start with a low-risk, high-volume process.
Publish Agent Cards for your most mature agents. Even if no one is consuming them yet, defining your agents’ capabilities in a structured format forces you to think about authentication, access control, and capability boundaries. That thinking is valuable regardless of protocol adoption.
The companies that will benefit most from the agent web are the ones building agents with interoperability in mind from day one, not the ones scrambling to retrofit it later.
Frequently Asked Questions
What is the agent web?
The agent web is an emerging network where AI agents from different companies discover, authenticate, and collaborate with each other using standardized protocols. It is analogous to the world wide web, which connected documents across servers. The agent web connects AI agents across organizational boundaries using protocols like Google’s A2A and Anthropic’s MCP.
How do cross-company AI agents communicate with each other?
Cross-company AI agents communicate using the A2A (Agent-to-Agent) protocol, launched by Google in April 2025 with over 50 enterprise partners. A2A enables task delegation, streaming responses, multi-turn conversations, and multimodal content exchange between agents from different vendors and organizations.
What is the difference between MCP and A2A protocols?
MCP (Model Context Protocol) connects AI agents to tools and data sources, like a universal adapter for external resources. A2A (Agent-to-Agent) connects agents to other agents for cross-organizational collaboration. They are complementary: MCP handles agent-to-tool communication, A2A handles agent-to-agent communication.
What are Agent Cards and how do they work?
Agent Cards are JSON metadata files hosted at /.well-known/agent.json that describe an agent’s capabilities, authentication requirements, supported formats, and endpoint URLs. They function like robots.txt for AI agents, enabling automated discovery. An orchestrator agent can crawl Agent Cards to find and evaluate available agents without hardcoded integrations.
Which companies are building cross-company agent networks?
Major enterprise vendors building cross-company agent capabilities include Salesforce (Agentforce), SAP (Joule on SAP Business Network), ServiceNow, Workday, Microsoft, Atlassian, and PayPal. Over 50 companies joined Google’s A2A protocol at launch, and both MCP and A2A are now governed by the Linux Foundation’s Agentic AI Foundation.