Debasree Panda 
We Have Seen This Movie Before
Every few years, enterprise technology gives us a new version of the same sentence:
“Why don’t we just build it ourselves?”
We have said it about Kubernetes platforms.
We have said it about observability.
We have said it about service meshes.
Now we are saying it about Agentic AI.
And honestly, I understand why.
The first version looks easy.
Connect an LLM. Give it tools. Add some automation. Put a nice UI on top.
Congratulations.
You have an agent.
Then someone asks:
Can it troubleshoot our production infrastructure?
And that’s when the fun begins.
Because enterprise Agentic AI isn’t just about the model.
It is about everything around the model.
The ARIA Framework: 7 Things to Think About Before You Build
At IMESH, we look at this through seven dimensions.
1. Architecture
Agentic AI isn’t frontend development with an LLM attached.
An enterprise agent has to deal with APIs, tools, identity, policies, data, networking, security and observability.
The agent may be intelligent.
But the infrastructure around it still has to work.
2. Talent
There are plenty of developers who can connect an application to an LLM.
There are far fewer engineers who understand:
AI + networking + security + distributed systems + observability.
And that’s exactly the combination production Agentic AI requires.
Finding those people is hard.
Keeping them is harder.
3. Build Is Easy. Maintain Is the Real Job.
Your POC works.
Everyone is happy.
Then six months later:
The model changed.
The framework changed.
Your APIs changed.
Your infrastructure changed.
Your security policies changed.
And suddenly your “AI engineer” needs an engineering team of its own.
4. AI Technology Changes Crazy Fast
This is perhaps the biggest problem with building everything yourself.
Models change.
Agent frameworks change.
MCP changes.
Routing changes.
Inference changes.
Pricing changes.
Your architecture may take 12 months to build around technology that changes every three months.
That isn’t future-proofing.
That’s future-proofing yourself into technical debt.
5. The Network Is the Real Complexity
This is the part I believe the AI industry often underestimates.
Agents don’t operate in isolation.
They communicate with:
- APIs
- databases
- services
- MCP servers
- other agents
- AI models
- gateways
- enterprise systems
Which means they depend on the network.
Routing matters.
Identity matters.
Authorization matters.
Security matters.
Observability matters.
Latency matters.
The model may be the brain.
The network is the nervous system.
And you don’t want an intelligent brain with a broken nervous system.
6. Context Is Everything
Ask a generic AI:
“Why is my service failing?”
You’ll probably get five plausible answers.
Ask an infrastructure-aware agent that understands your environment, topology, policies, configuration and telemetry?
Now you’re having a different conversation.
An enterprise agent doesn’t just need intelligence.
It needs context.
Otherwise, it’s essentially a very confident intern.
7. The Economics Change at Scale
Your POC might make 100 model calls.
Production might make millions.
Your POC might have one cluster.
Production might have dozens.
Your POC might need one engineer.
Production needs security, platform engineering, SRE and operations.
So, the real question isn’t:
“Can we build it cheaper?”
It’s:
“Can we operate it better and cheaper for the next five years?”
That’s the Build vs. Buy question that actually matters.
So, Should Enterprises Build or Buy?
I don’t believe everything should be bought.
And I don’t believe everything should be built.
Build what differentiates your business.
Platform what doesn’t.
If your proprietary AI workflow is your competitive advantage, build it.
But should every enterprise also build its own infrastructure intelligence layer, networking intelligence, observability, security controls, agent lifecycle and troubleshooting system?
I’m not convinced.
We didn’t all build our own databases.
We didn’t all build our own networking stacks.
And we probably shouldn’t all build our own AI infrastructure engineer.
This Is Why We Built ARIA
At IMESH, we call it:
ARIA – Autonomous Reliability & Infrastructure Agent
The idea is simple:
One Agent. Your Entire Cloud-Native Infrastructure.
ARIA is designed to work across the infrastructure technologies enterprises already rely on including Istio, Envoy Gateway, Agent Router (EAIG) and Cilium.
It is built around the operational problems that become increasingly difficult as infrastructure becomes more distributed:
- Architecture
- Security
- Observability
- Troubleshooting
- Optimization
- Lifecycle upgrades
Because we don’t believe the future is simply about having more AI agents.
We believe the future is about having AI that actually understands the infrastructure those agents run through.
The Final Test
Before starting your next internal Agentic AI project, ask yourself:
Can we build it?
Probably.
Can we maintain it?
That’s the real question.
Can we secure it?
Can we give it the right context?
Can it understand our network?
Can we keep it current as AI changes?
Can we afford to own it for five years?
If the answer is yes to all seven:
Build it.
If not, don’t underestimate what you’re signing up for.
Because the most expensive sentence in enterprise technology is still:
It shouldn’t be that difficult.
It usually is.
And that’s exactly why we built ARIA.
