Featured

Episode-XXVIII: The AI Engine Is Ready. But Where’s The Rest?

Authors: Tushar Katarki , Fatih Nar, William Caban

The Engine to Wheels Icon

The evolution of generative AI mirrors the combustion engine’s journey from novelty to necessity. Early engines were simple, single-cylinder affairs, much like our first AI models. Then came the race for power; more cylinders, larger displacement, higher compression. Today’s massive language models follow the same trajectory, growing from millions to hundreds of billions of parameters (and now even trillion parameters).

But here’s what most executives miss/ignore:

A powerful engine without a gearbox and transmission system is just expensive noise (i.e. just yet another chatbot)!

The Engine-to-Wheels Problem

Consider your car’s drivetrain. The engine generates power, but it’s the gearbox that makes that power useful shifting between torque for climbing hills and speed for highways. The transmission then carries this optimized power to where it matters; the wheels that actually gets you going to your destination!

Figure-1 Engine-to-Wheel Diagram

However, even the best transmission requires transmission fluid, the often-overlooked component that enables smooth gear changes, reduces friction, and prevents the entire system from grinding to a halt.

In the AI world, we’ve been obsessed with building bigger engines (models) while neglecting the critical systems that deliver intelligence where value is needed to be created. This is why many AI initiatives stall despite massive computational investments.

The new reality:

• AI agents are becoming our gearboxes, intelligently shifting between different “gears” of computational intensity <-> model capacities & capabilities.
• Network infrastructure acts as the transmission, carrying intelligence to every corner of your business operations over distributed geographies.
• Protocols like Model Context Protocol (MCP) and Agent to Agent Communication Protocol (A2A) serve as the transmission fluid enabling smooth communication & inter-working between AI components, facilitating agent-to-agent collaboration, and ensuring your AI systems work together seamlessly and in an optimized harmony.

The Critical Role of AI “Transmission Fluid”

Just as transmission fluid enables smooth gear shifts and power transfer, modern AI protocols are becoming essential for distributed AI architectures. They come in three critical forms, with the Llama Stack serving as a foundational layer:

Llama Stack: The Foundation

The Llama Stack is an emerging open-source ecosystem that provides critical platform components for distributed AI architectures. The Llama Stack provides the building blocks and design patterns for creating MCP & A2A protocols, as well as the applications that leverage them. By offering a common foundation, the Llama Stack accelerates the development and adoption of these crucial components in distributed AI systems.

Reach for Comprehensive Abilities with Model Context Protocol (MCP)

• Able to Leverage Bigger Engines: Remote access to bigger/better models for offloading extensive tasks where there are no/less-resource restraints.
• Enable Seamless Tool Access: AI models can smoothly interact with databases, APIs, and all necessary supporting enterprise ecosystems.
• Reduce Integration Friction: Standardized communication means less custom connector development.
• Accelerate Deployment: Pre-built protocol support speeds time-to-value.

Agent-to-Agent Protocol (A2A)

• Enable Multi-Agent Orchestration: Agents can negotiate, collaborate, and hand off tasks seamlessly among themselves.
• Create Emergent Intelligence: The sum of multiple specialized agents working together exceeds the individual value/intelligence of an agent.
• Build Resilient Systems: Agents can route around failures and optimize workflows dynamically.

When built on top of the Llama Stack framework, the MCP and A2A protocols act as the “transmission fluid” that ensures smooth operation and collaboration within distributed AI architectures. As these protocols mature and gain adoption, they will play a crucial role in enabling the development of sophisticated multi-agent systems that can tackle complex real-world challenges.

Figure-2 Stacks Architecture (Reference; Episode-XXVII Lessons learned from MCP Experiments)

Without the Llama Stack as a foundation and the MCP and A2A protocols as the communication layer, you might have a powerful engine and a sophisticated transmission. Still, the gears would grind, agents would work in silos, and intelligence wouldn’t flow where it’s needed. By providing a common ground for building these critical components, the Llama Stack accelerates the journey towards truly distributed and collaborative AI systems.

The Business Case for Distributed & Resilient Intelligence

Cost Optimization Through Edge Processing

• Short-term opportunity: Service providers can deploy lightweight agents at edge data centers for network optimization tasks. These agents use A2A to coordinate with regional and central systems only when necessary, dramatically reducing backhaul costs while improving response times. Think real-time traffic management, local content caching decisions, or automated fault detection all handled at the edge with occasional synchronization.
• Mid-term potential: As agent protocols mature, expect sophisticated multi-agent systems where access agents (say radio access network agents) collaborate with core network agents for dynamic resource allocation. A security agent might coordinate with a capacity planning agent to prevent distributed denial of service (DDoS) attacks while optimizing spectrum efficiency, all without central intervention. The key is standardized communication protocols that make this collaboration seamless.
• Long-term vision: Self-organizing agent networks that dynamically form task forces, share learnings, and evolve strategies (example; autonomous networks in our Telco-AIX repo). Imagine 5G networks where slicing agents, QoS agents, and customer experience agents continuously optimize operations through local collaboration, only escalating to the cloud for strategic decisions.

New Revenue Streams via Embedded Intelligence

The immediate opportunity for telcos lies in deploying provider-aware specialized agents (agents that understand their business and operations) that enhance existing services. For example, network slices embedded with performance optimization agents don’t just guarantee SLAs; they can negotiate with enterprise customer agents for dynamic resource allocation while coordinating with billing agents to implement real-time pricing models.

As protocols standardize, we’ll see agent marketplaces emerge (example; Linux Foundation Projects Essedum & Agntcy). Your network optimization agent could deliver insights to MVNO partners. Your customer experience agent could offer predictive churn services to enterprise clients. Network operators transform from connectivity providers to intelligence service providers.

The long-term vision extends to economic ecosystems where agents autonomously create value through complex negotiations and collaborations. Picture smart cities where service provider infrastructure agents bid on IoT connectivity tasks, coordinate with municipal service agents, and optimize resource usage across multiple verticals, all enabled by robust A2A protocol.

Competitive Advantage Through Agent Collaboration

In complex environments like multi-vendor service provider networks or global roaming partnerships, agent teams already demonstrate their ability to respond faster than monolithic systems. A congestion detection agent spots an anomaly, alerts a traffic management agent, which coordinates with a slice orchestration agent, all in milliseconds. Operators who master this agent choreography consistently outperform those relying on traditional network management systems.

As A2A protocol standardizes, the real opportunity emerges in cross-carrier agent collaboration. Your roaming agents negotiate directly with partner network agents, settlement agents, and quality assurance agents to optimize entire service chains. The competitive advantage shifts from having the best individual network to orchestrating the best agent ecosystems.

The Infrastructure Reality Check

As discussed in the recent article “Episode-XXVI: The AI Rings”, organizations must strategically deploy AI models across different tiers to maximize their potential while minimizing risks and dependencies. This approach aligns with the concept of distributed AI architecture and the need for a robust infrastructure to support the collaboration of intelligent agents.

Building distributed AI capability requires a strategic assessment of technology maturity across three layers:

(1) Hybrid Model Approach for Accuracy

While smaller models can be improved with fine-tuning and inference time scaling, accuracy remains a critical challenge, especially for semi-autonomous and autonomous network operations. A hybrid approach that leverages very large models for complex network planning and optimization reasoning, combined with smaller models for executing specific tasks like anomaly detection or traffic prediction, might be necessary to ensure the required accuracy levels.

This hybrid approach balances the strengths of large models in complex reasoning with the efficiency and specificity of smaller models for targeted execution. By distributing the workload intelligently, service providers can optimize for both accuracy and performance while meeting stringent latency requirements.

(2) Network as Neural Pathways

Your service network transforms from simple data pipes to the nervous system of your AI strategy. 5G and private networks enable ultra-low latency agent communication. Network slicing allows agents to have guaranteed communication channels. When connections fail, redundant pathways ensure agent networks self-heal and continue operating a concept telcos understand better than any other industry.

(3) Edge Compute as Local Intelligence

Edge infrastructure evolves from simple hosts to sophisticated agent platforms:

• RAN intelligent controllers run agents that coordinate beam forming, power optimization, and interference management.
• Customer premises equipment hosts collaborative agents for monitoring and troubleshooting service quality.
• Mobile edge computing nodes deploy agents that negotiate workload distribution and predict capacity needs.
• Network functions operate as intelligent agents that optimize traffic flows in real-time.

Protocols as the Lubricant

The binding force comes from robust protocols. MCP enables seamless OSS/BSS integration, allowing agents to access any data source or service. A2A facilitates rich agent interactions, from simple handoffs to complex negotiations between network domains. Security protocols ensure that all communications remain private and authenticated, which is critical for maintaining network integrity. Without these protocols, your distributed AI fragments into isolated islands of intelligence.

Service providers must adopt off-the-shelf models, along with retrieval augmented generation (RAG), purpose-built system prompts, and output alignments, not only for optimal outcome generation but also to comply with stringent regulatory requirements.

Build, Buy, or Partner: The Strategic Choice

• Build when distributed AI represents your core competitive differentiator. If you’re pioneering new network automation models, developing proprietary service orchestration strategies, or creating unique customer experience optimizations, building makes sense. Leading operators creating novel 5G monetization strategies or developing autonomous network capabilities often choose this path.
• Buy proven solutions for standard use cases. Agent frameworks with protocol support, pre-built libraries for common network operations, and established multi-agent platforms can accelerate your journey. Focus your innovation on service differentiation rather than reinventing foundational capabilities.
• Partner for ecosystem participation. No operator can create all the protocols, standards, and governance frameworks alone. Partner through industry bodies like TM Forum, GSMA, or O-RAN Alliance for protocol development, agent marketplaces, cross-carrier networks, and security frameworks. The winners will balance internal innovation with ecosystem leverage.

The ROI Evolution

Distributed AI fundamentally changes ROI calculations:

Traditional AI ROI = (Value Created — Cloud Costs) / Investment

Agent-Enabled AI ROI = (Value Created + Edge Savings + Agent Synergies + Network Effects — Infrastructure) / Investment

The promise of distributed AI lies in multiple value drivers working together. Edge processing reduces backhaul and cloud compute costs by keeping network operations local. Response time improvements enable the introduction of new low-latency services that were previously impossible with centralized processing. Agent collaboration creates operational efficiencies that no single AI could achieve. And as your agent network grows, network effects multiply value exponentially.

While specific results vary by operator and implementation, organizations consistently report that the combination of OPEX reduction, service quality improvement, and new revenue capabilities creates compelling returns. The key insight is that infrastructure investments amortize across multiple use cases, while agent capabilities compound over time.

The key insight: Infrastructure investments amortize across multiple use cases, while agent capabilities compound over time.

Action Plan

• Phase-1 [Assessment] : Audit your network operations and customer service workloads to identify tasks suitable for edge deployment and agent collaboration. Look for high-frequency operations like traffic optimization, location-specific services, or latency-sensitive applications. Map current OSS/BSS integration challenges that protocols could solve.
• Phase-2 [Protocol Evaluation] : Assess readiness for MCP and A2A protocols across your network domains. Identify which systems need upgrades from RAN to core to IT systems. Choose initial protocol standards that align with your existing architecture. Remember: protocol decisions made now determine future scalability.
• Phase-3 [Pilot Development] : Launch a focused pilot with 2–3 cooperating agents. Start simple, perhaps network optimization and customer experience agents working together in a specific region. Implement both MCP for tool access and A2A for agent coordination. Measure collaboration benefits rigorously against traditional approaches.
• Phase-4 [Scale Planning] : Based on pilot results, design your broader agent ecosystem. Identify additional use cases across network planning, operations, and customer service. Plan required infrastructure upgrades and partnership opportunities. Build your implementation coalition across network operations, IT, business units, and technology partners.

Future Look

As we stand at a break point in-time , just as automobiles evolved from simple engines to sophisticated systems with transmissions, computer controls, and integrated components; service networks must also evolve from isolated systems to collaborative agent networks.

The executives who grasp this shift will build networking systems where agents collaborate naturally, MCP enables frictionless OSS/BSS integration, A2A protocol facilitates rich interactions between network domains, and edge and cloud agents work in perfect concert. They’ll create self-organizing networks that learn and adapt collectively, generating value that compounds over time.

The journey unfolds in stages:

• Today: Deploy basic agents for network optimization and customer service automation.
• Tomorrow: Build sophisticated multi-agent systems that self-organize across network domains.
• The Day After Tomorrow (Hopefully Not a Doomsday 🥶): Participate in global agent economies creating unprecedented value through autonomous networks.

The engine of AI is ready -> The transmission is being built -> The protocols; both MCP and A2A are emerging as the vital fluid that makes it all work. The question is:

Will you build the complete AI drivetrain that transforms raw intelligence into collaborative, distributed advantage across your entire network?

The race isn’t about having the most powerful AI -> it’s about orchestrating intelligent agents that work together seamlessly, communicate efficiently, and deliver value at every level of your network. That race has already begun!

The journey from centralized AI to distributed agent networks requires vision, strategy, and commitment to open source platforms & protocols. Service providers that start building their agent ecosystems today, complete with robust communication protocols, will define how networks operate and services are delivered tomorrow.