The new challenge for networks: autonomous AI agents

Over the past two years, artificial intelligence has entered a new era. We are no longer talking about chatbots and text generators — agentic AI systems have emerged that autonomously browse, code, call APIs, make decisions, and communicate with other AI agents without human intervention. This shift is fundamentally reshaping enterprise network traffic patterns, security requirements, and infrastructure needs.

Explosive traffic growth

According to the Nokia Global Network Traffic Report, enterprise AI traffic is growing at a CAGR of 57% and is projected to reach 2,192 exabytes per month by 2033. Akamai research shows AI bot traffic has surged over 300% in the past two years, while Cloudflare data indicates AI bots accounted for an average of 4.2% of HTML requests in 2025.

Why is the difference between traditional and agentic traffic so dramatic? The reason is simple: a single user request can generate 10 to 50 API calls as the AI agent decomposes tasks, invokes tools, retrieves data, and coordinates with other agents. This cascading, chained traffic represents an entirely new type of network load.

East-west traffic: the new challenge

Traditional network architectures were built for north-south traffic — users send requests to servers, servers respond. AI agents, however, primarily generate east-west traffic: machines communicating with machines, servers with servers, APIs with APIs. According to Akamai, more than 76% of modern data center traffic now flows east-west.

Traditional hub-and-spoke and MPLS architectures were not designed for this. Agent-to-agent communication requires sub-50 millisecond response times — even 100ms of latency can break reasoning chains and disrupt action sequences.

Security risks: the OpenClaw lesson

Agentic AI brings not only traffic challenges but serious security risks. Palo Alto Networks Unit 42 warns that agentic applications inherit vulnerabilities from both LLMs and external tools while expanding the attack surface through autonomous decision-making and dynamic tool invocation.

The OpenClaw case illustrates the dangers well: over 40,000 exposed instances of the AI agent system were found within 24 hours, 63% were exploitable, and more than 1,800 instances were leaking API keys and credentials. The case demonstrates that even popular, well-known tools can open serious security gaps.

OWASP-identified agentic AI risks

Agent Goal Hijack — manipulating an agent's objectives
Tool Misuse — using tools in unintended ways
Identity and Privilege Abuse — agents escalating their own permissions
Memory Poisoning — corrupting agent long-term memory to alter future behavior
Supply Chain Vulnerabilities — compromised tools in the agent's toolchain

Identity management: traditional systems are inadequate

According to the Cloud Security Alliance, traditional IAM systems (OAuth, OIDC, SAML) are fundamentally inadequate for managing AI agents. Agents are ephemeral — they spin up and shut down dynamically. They assume multiple roles, require context-aware credentials, and machine-to-machine communication does not fit human authentication models.

New frameworks are emerging: Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs) for agent identity. HashiCorp is developing zero trust solutions for agentic systems, focusing on managing non-human identities at scale.

The infrastructure readiness gap

The Cisco AI Readiness Index 2025 reveals alarming data:

Only 28% of organizations believe their infrastructure can handle AI workloads
Only 15% have networks fully ready for AI
54% report their networks cannot scale for current complexity
Gartner predicts 40% of enterprise applications will feature task-specific AI agents by 2026

New protocols: MCP and A2A

Two new protocols are shaping the future of agentic networking:

MCP (Model Context Protocol) is Anthropic's open-source standard providing a unified connection between AI applications and external systems. MCP gateways serve as a single checkpoint for all agent-tool interactions.

A2A (Agent-to-Agent Protocol) is Google's protocol enabling agents from different frameworks to collaborate. A2A uses "Agent Cards" for discovery — comparable to DNS for agents.

As Salt Security warns: "The widespread adoption of MCP and A2A will inevitably lead to more APIs and greater API usage, not less. This explosion of internal traffic is the perfect hiding place for attackers."

Firewalls and traffic shaping: rethinking required

Traditional firewalls are built on perimeter defense, but agentic east-west traffic entirely bypasses traditional firewalls as it stays within the data center. The solution: microsegmentation embedded directly into data center switches.

Traditional rate limiting is also inadequate: AI agents behave like legitimate high-volume consumers but share characteristics with malicious botnets. According to Nordic APIs, Adaptive Rate Limiting (ARL) is needed — dynamically adjusting limits in real time based on context and behavioral analysis.

What can companies do?

ITEX Solutions recommends the following steps for any company deploying or planning to deploy AI agents:

Bandwidth audit — Even small AI agent deployments generate 10-100x API traffic compared to traditional workflows. Plan for at least 2-3x growth.
API gateway investment — Without API gateways with adaptive rate limiting, you risk either blocking your own agents or leaving APIs open to abuse.
Security audit — Check all AI agent deployments for exposed endpoints, leaked credentials, and overly permissive access.
Zero Trust implementation — Every AI agent must be treated as an untrusted entity with machine identity management and just-in-time access controls.
Monitoring upgrade — Traditional monitoring is blind to AI agent behavioral anomalies. Invest in agent-specific observability systems.

Conclusion

Agentic AI is not a future vision — it's already here. According to Gartner, 33% of enterprise software will include agentic AI by 2028. Companies that prepare their networks and security systems now will gain a significant competitive advantage. Those that don't may face the 40% project cancellation rate that Gartner predicts.

The ITEX Solutions team is ready to help your company prepare its network infrastructure for the agentic AI era. Contact us for a comprehensive network and security audit: info@itex.hu | +36 20 615 5191

Networking Challenges in the Agentic AI Era