By Santosh Dornal, VP of Engineering, Alkira
Production AI agents rarely fail in obvious ways. There is no dramatic crash, no clear exception, no instant signal that something has broken. More often, performance degrades quietly. Responses become partial. Latency stretches. Outputs look plausible but arrive missing key context. When teams investigate, they often find that the issue is not the model itself, but a missing connection, an access gap, or an unreliable path between systems.
Much of the industry conversation around AI agents still centers on model choice, prompt engineering, or retrieval frameworks. Those things matter, but they are only part of the picture. Once agents move from isolated demos into production systems, a different set of issues becomes more important: can the agent reliably reach the tools it depends on, maintain access across environments, and operate predictably when something in that chain degrades?
AI Agents are Distributed Systems
One of our early lessons at Alkira was that agents cannot be treated like advanced chatbots. In production, an agent only becomes useful when it can pull together data from multiple systems, whether that means checking Kubernetes clusters, reading metrics and logs, or interacting with internal tools and workflows.
In a demo, those dependencies are usually close together and tightly controlled. In a real enterprise environment, they are distributed across clouds, security zones, platforms, and operational domains. At that point, the quality of the agent depends not just on the model, but on the reliability of the network paths connecting everything around it. When those paths become inconsistent, agents do not always fail cleanly. They stall, retry, time out, or return incomplete answers with enough confidence to look correct. That is what makes these failures difficult to diagnose and risky to ignore.
The Economics of Continuous Agents
Cost becomes a real design constraint once agents move into continuous operation. Many are not invoked occasionally; they run all the time, observing systems, correlating signals, and supporting background workflows. In that model, inference costs can scale quickly if every decision depends on repeated calls to a large external model.
We learned early that usefulness alone is not enough. An agent also has to be economically sustainable. That means separating lightweight observation from deeper reasoning, being intentional about when inference is actually needed, and choosing an execution model that balances flexibility, cost, and control. The larger lesson is simple: in production, AI architecture has to be designed as an operating system concern, not just a model selection exercise.
Solving the Access Problem with Controlled Gateways
Access is another place where production reality changes the problem. Agents need credentials, APIs, and permissions to do useful work, but broad access without strong controls creates long-term risk. Security and connectivity have to be designed together. If the access layer is too open, the risk is obvious. If it is too brittle, the agent becomes unreliable.
Our approach was to use Model Context Protocol servers as controlled gateways, defining explicitly what an agent could see, what it could call, and what actions it could take. We started with a strict read-only model and focused first on observation, traceability, and auditability. That made it easier to build trust while reducing the chance of unintended consequences in live environments. In practice, teams tend to have more success with AI agents when they design this infrastructure and permission layer early, rather than treating it as something to bolt on later.
Connectivity Is Part of the Intelligence Layer
One of the clearest lessons from our work is that reliable AI agents depend on reliable infrastructure. Stable connectivity is not just an operational prerequisite, but it directly affects the quality of what the agent can observe, correlate, and return. Without a predictable, secure network foundation, agents become inconsistent interpreters of system state.
That is why the infrastructure layer deserves more attention in the AI agent conversation. Not because it is more visible than the model layer, but because it is often where production systems ultimately succeed or fail. In our experience, practical AI starts with dependable day-to-day operations: controlled access, stable inference paths, and a network foundation robust enough to support the intelligence layered on top of it. In enterprise environments, that is not background plumbing. It is part of what makes agentic systems viable in the real world.
Agents need stable connectivity across environments, secure access to the right tools and data, and an execution model that can hold up under real operational demands. As more enterprises move from experimentation to live deployment, the vendors that invest in this foundation will be better positioned to deliver AI that is not only intelligent, but reliable, secure, and operationally sound.
##
ABOUT THE AUTHOR
Santosh Dornal is the VP of Engineering, Infrastructure & Operations at Alkira, where he leads the teams responsible for scaling cloud-native networking and DevOps. With over 15 years of experience in the networking industry, Santosh has been a key architectural driver behind large-scale SD-WAN, mobile gateway, and cloud connectivity platforms.
Before joining Alkira, he was a member of the original team at Viptela, where he focused on solving complex cloud connectivity challenges for global enterprise deployments. His career also includes engineering leadership roles at Cisco (via the Starent Networks acquisition), Juniper Networks, and Ixia. Santosh holds an MS in Computer Networking from I2IT Pune and a B.Tech in Electronics and Communications.
