Skip to content
MAF Learning Hub
Advanced Duration: 25 min

Managing network topographies and egress control

Prerequisites: automating-infrastructure-iac

Agents call LLMs, tools, and each other - which means they reach across network boundaries constantly. This lesson covers designing the network topology for agent workloads and, critically, egress control: deciding where agents are allowed to send data.

What you will learn

  • How to think about network topology for agent workloads.
  • What egress control is and why it is essential for agents.
  • How the Omni Gateway acts as the enforcement chokepoint.

Network topology for agents

A sound topology isolates workloads (for example, separating production agents from experimental ones) and defines the paths traffic may take between runtimes, gateways, and external endpoints. Isolation limits blast radius when something goes wrong.

Egress control

Egress control restricts outbound connections - which external LLMs, MCP servers, or APIs an agent may reach. For autonomous agents this is a primary safety control: it prevents data exfiltration and stops an agent from calling unapproved endpoints.

# egress-policy.yaml (excerpt, illustrative)
egress:
  default: deny
  allow:
    - host: api.approved-llm.com
    - host: mcp.internal.example.com

Default deny

Start from 'deny all egress' and allow-list only the endpoints agents genuinely need. This is Zero-Trust applied to the agent's outbound calls.

The Omni Gateway as chokepoint

Routing agent traffic through the Omni Gateway gives you a single enforcement point for these rules, so egress decisions are consistent and auditable rather than scattered across workloads (Securing Agent Interactions with Omni Gateway). (Forward-looking: exact egress-policy configuration is illustrative; confirm against official documentation.)

Next steps

Continue to Cost management and governance strategies to keep the platform economical as agent usage scales.

References