Technology for the Agentic Age: Why Infrastructure Determines Outcome

In Part 1 of this series, I introduced the People-Process-Technology framework for TMT's agentic transformation. Technology may not be first in that sequence, but it is first in consequence. Without the right infrastructure, even the best processes and people hit a wall.

The gap lies in the architecture itself.

Most TMT organizations have invested heavily in cloud migration and data modernization. Yet 74% of enterprises still lack the foundational capabilities required to move beyond proof-of-concept to production-scale AI solutions.

Meanwhile, 70% of top performers report difficulties integrating data into AI models, citing challenges across data quality, governance processes, and training data availability.

How do we design scalable, cost-conscious, and secure tech stacks that can adapt as models and tools evolve?

This post answers that question across four dimensions: foundational data imperatives; agent orchestration; observability and cost management; and sector-specific architecture requirements.

Foundational Data Imperatives

Agentic AI is only as capable as the data it can access. For TMT enterprises, this means unifying fragmented OSS/BSS systems, content management platforms, and customer data stores into agent-accessible architectures.

Real-time data accessibility. Traditional batch processing won't cut it. Agents need event-driven pipelines they can query in milliseconds. The shift from scheduled ETL jobs to streaming architectures has become a prerequisite.

Unified data models. Service assurance, revenue management, subscriber 360 views—these domains can't live in silos if agents need to coordinate across them. Consider what a telecom-specific unified model requires: 

Domain models spanning Care, Customer, Network, Sales, Finance, and Marketing—each with its own dimensional structure, fact tables, and aggregations. The Network domain alone might include multiple dimensions and fact tables covering network strength, speed, site location, and call traffic. The Customer domain integrates 360-degree profiles, loyalty data, preference signals, and segmentation. 

When these domains connect, agentic use cases become possible: precision network deployment informed by customer density, revenue recognition models that link transactions to subscriber behavior, and disconnect prediction that correlates care interactions with churn signals.

The goal: a single pane of glass for multi-domain agent coordination.

Structured and unstructured integration. Agents need access to legacy contracts, technical blueprints, call transcripts—all of it. This is where retrieval-augmented generation (RAG) architectures become essential, supported by vector databases and semantic search capabilities. First-party and third-party data access over secured Model Context Protocol (MCP) servers enables agents to pull from diverse sources while maintaining governance boundaries.

TMT-specific scale considerations. Tier-1 operators now ingest upwards of five billion KPI samples daily, ten times the volume from just five years prior. Media companies face similar explosions in content metadata and behavioral signals. Architecture must accommodate this volume without sacrificing the latency agents require.

Agent Orchestration and Communication

Individual agents solving individual problems won't scale. TMT enterprises need orchestration frameworks that enable multi-agent collaboration, cross-domain coordination, and human-in-the-loop governance.

The orchestration layer. At minimum, this includes agent registry and discovery services, shared memory and context management, policy enforcement and guardrail frameworks, and tool-calling and API orchestration. Without standardization, you end up with fragmented deployments that can't communicate, which is exactly the tool sprawl problem I outlined in Part 1.

Communication protocols: MCP vs. A2A vs. Universal Commerce Protocol (UCP). Three emerging standards are shaping how agents connect to their environment and each other.

Why does this matter for TMT? A telecom network optimization agent should be able to coordinate with a customer service agent without custom integrations. A media company's recommendation engine should connect seamlessly to commerce fulfillment. Standardized protocols reduce brittleness and accelerate deployment.

Multi-agent patterns. Depending on the use case, TMT enterprises will deploy hierarchical structures (supervisor agents coordinating specialists), peer-to-peer collaboration, or hybrid approaches. The key is establishing a common orchestration methodology that gives developers flexibility to use frameworks like LangGraph, CrewAI, or Semantic Kernel while maintaining enterprise-wide consistency.

Agentic Commerce. Commerce infrastructure deserves specific attention because it sits at the intersection of TMT's customer-facing operations. Commerce systems must expose capabilities through standardized protocols (UCP, MCP, A2A) so agents can discover offerings, negotiate transactions, and complete purchases autonomously, all while maintaining audit trails and policy compliance.

Authentication and security. Agents acting autonomously require a robust auth ecosystem based on enterprise guidelines. This means implementing common authentication principles across both agents and MCP server-based data access. This ensures that autonomous actions remain auditable and policy-compliant.

Observability and Cost Management

Agentic systems introduce operational complexity that traditional monitoring wasn't built to handle. Without built-in observability and FinOps discipline, costs tend to spiral and failures compound.

AgentOps and observability. You need standardized tooling to monitor agent thought processes and iterations. This includes the ability to evaluate agent output synchronously, for real-time validation, and asynchronously, for batch analysis. Reusable components for hallucination detection, toxicity detection, and bias monitoring should be built once and deployed across the agent ecosystem.

The goal is an LLMOps observability dashboard providing complete visibility across deployment patterns and agentic workflows. Without this, debugging multi-agent failures becomes nearly impossible.

Cost challenges unique to agents. Multi-agent conversations generate exponentially more API calls than single-model deployments. Each reasoning step consumes tokens. Memory persistence across millions of customer interactions adds storage and compute costs. Continuous model retraining compounds the expense.

AI workloads defy traditional forecasting. Volatile usage patterns, fragmented cost structures across tokens and compute, and rapid model evolution that can swing performance and expenses dramatically. This represents a new discipline distinct, in many ways, from traditional cloud FinOps.

FinOps requirements for agentic AI. Token-level cost tracking by use case. Model routing optimization (i.e., matching task complexity to the right-sized model). Infrastructure autoscaling with cost constraints baked in. And critically, ROI measurement frameworks tied to business outcomes, not just technical metrics.

IDC warns that by 2027, G1000 organizations will face up to a 30% rise in underestimated AI infrastructure costs. The enterprises that build financial observability into their agent architectures from day one will avoid the budget shocks hitting their competitors.

Guardrails and responsible AI. Input/output scanning for harmful content. Hallucination detection and grounding validation. Policy enforcement for agent autonomy boundaries. Human-in-the-loop escalation triggers. These are now must-haves for production deployment.

Sector-Specific Architecture Requirements

While foundational principles apply across TMT, each sector has distinct infrastructure requirements based on data types, latency constraints, and integration landscapes.

Telecommunications. The primary challenge is legacy OSS/BSS integration with modern agent platforms. Priorities include network telemetry ingestion at scale via streaming architectures, digital twin integration for network simulation, and self-healing network capabilities.

Telco use cases already proving value: 

• AI agents that monitor and self-heal network issues during peak hours
• Customer service agents with real-time access to subscriber data and network status
• Autonomous capacity planning that adjusts infrastructure allocation based on predicted demand

Media and Entertainment. The primary challenge is real-time personalization across massive content catalogs. Priorities include behavioral signal processing (micro-behavioral triggers), content metadata enrichment and semantic search, rights management integration for multi-territory compliance, and dynamic content assembly pipelines.

Media and entertainment use cases gaining traction: 

• Agentic personalization that tailors recommendations based on viewing behavior in real time
• Automated content tagging and classification at scale
• AI-driven ad targeting that optimizes campaign performance without additional production spend

Technology and SaaS. The primary challenge is embedding agent capabilities into existing product architectures. Priorities include API gateway design for agent-to-agent communication, multi-tenant agent isolation and security, product telemetry integration for agent training, and customer data platform connectivity.

Tech & SaaS use cases driving competitive advantage: 

• Agent-driven onboarding assistants that handle customer setup autonomously
• Intelligent copilots embedded as core product features rather than add-ons
• Self-optimizing workflows that reduce churn and increase product adoption

The Infrastructure Imperative

Fewer than one in five companies report meaningful earnings impact from their gen AI investments. The gap between experimentation and enterprise-scale value isn't closing automatically. It requires deliberate architectural investment.

Technology infrastructure determines what's possible. Without real-time data accessibility, agents can't act on current information. Without orchestration standards, multi-agent coordination becomes a custom integration nightmare. Without observability and cost controls, budgets balloon before value materializes. Without sector-specific considerations, generic deployments underperform.

The TMT enterprises that get infrastructure right will not only deploy agents, but scale them. And scale is where the Agentic Age separates leaders from laggards.

In Part 3, we'll examine how to redesign processes for agent-first operations, because bolting AI onto legacy workflows won't capture the value that purpose-built processes unlock.

AUTHOR - FOLLOW
Anugraha Sinha
Director, Generative AI

AUTHOR - FOLLOW
Shree Damani
Senior Director & Head of Sales - TMT, Tredence Inc.