Building Agentic AI and Autonomy at Global Scale

Prince Victor Orjiugo
Sep 28
8 min read

The relentless march of artificial intelligence continues to reshape industries, moving beyond predictive analytics and into a new era of autonomous action. At a recent Stockholm AI summit, three distinct but interconnected presentations painted a vivid picture of this evolution, offering deep insights into the architecture of autonomous networks, the practicalities of building enterprise-grade AI agents, and the strategic vision of a frontier AI company.

This post synthesizes the core ideas from the first three speakers: Jörg Niemöller of Ericsson, who demystified the transition to autonomous networks; Corentin Petit from Mistral, an Enterprise AI specialist who detailed the challenges and triumphs of deploying AI agents in large corporations; and Christian Ryan from Anthropic, who provided a look into the philosophical and technical underpinnings of building effective AI agents.

Part 1: The Intent-Driven Future of Networks – An Ericsson Perspective

The foundation of our digital world is the vast, intricate web of telecommunication networks. For decades, managing these networks has been a profoundly human endeavor. However, as Jörg Niemöller from Ericsson articulated, we are at a pivotal moment in the evolution of network operations, transitioning from manual control to a future of autonomous, intent-driven systems.

The Evolution from Manual to Autonomous

Niemöller laid out a clear evolutionary path for network operations, categorized into three distinct stages:

Manual Operations: The traditional model, where human operations teams are responsible for every step of the process: observing network performance, analyzing data, making decisions, and manually executing changes. This is a labor-intensive, slow, and often reactive process.
Automated Operations: The current state for many advanced networks, this stage introduces automation for defined static policies and workflows. While humans are still in the loop for decision-making, the execution of those decisions is increasingly handled by machines. This brings significant efficiency gains but still relies on human oversight and predefined rules.
Autonomous Operations: The final frontier. In this paradigm, the system itself makes decisions autonomously based on high-level intent. Humans set the goals and constraints, but the network itself figures out the how. Machine decision-making and machine execution become the norm, with humans moving into a supervisory and strategic role.

The Five Levels of Autonomy

To further clarify this journey, Niemöller introduced a framework of five levels of autonomy, analogous to the levels of autonomous driving. This framework isn't just about technology; it's about the shifting relationship between people and systems.

Level 0 & 1 (Manual & Assisted Operation): The system is primarily operated by people. AI might provide some assistance, but humans are firmly in control of all aspects of operation and maintenance.
Level 2 & 3 (Partial & Conditional Automation): This is where we see a significant shift. The system becomes an active partner, handling more of the execution, awareness, and analysis. People and systems work in tandem, with the system taking on more responsibility under specific conditions.
Level 4 & 5 (High & Full Automation): At these levels, the system is in control. It's capable of making its own decisions and taking action to achieve the goals set by its human operators. This is the ultimate vision of an autonomous network.

The Central Role of "Intent"

The key to unlocking these higher levels of autonomy is the concept of "intent." Instead of programming a network with a long list of specific rules, an intent-based system allows operators to declare their desired outcomes in a more abstract, business-oriented language.

Niemöller presented a compelling diagram illustrating this concept. An autonomous network sits at the center of a complex ecosystem of stakeholders, each with their own intents:

Service Providers: Their intent is driven by business goals like revenue assurance, cost optimization, quality of service, and security.
Customers: They have their own set of intents, which can be diverse and sometimes conflicting. One customer might prioritize low price, another might demand high performance for gaming, and a third might be most concerned with data privacy.

The autonomous network's job is to take all these competing intents as input, understand them, and then translate them into concrete actions that optimize the network's performance to meet these goals.

The Intent Management Function (IMF)

**Building Agentic AI and Autonomy at Global Scale**

To make this a reality, Ericsson uses the "Intent Management Function" (IMF) as the core building block of autonomous network operation. The IMF is a closed-loop system that continuously:

Receives Requirements: It takes in high-level intents from operators and customers.
Gathers Knowledge: It constantly measures the state of the network and uses its knowledge base to understand the current context.
Makes Decisions: It analyzes the intents and the network state to decide on the best course of action.
Takes Action: It executes the necessary changes in the network.
Reports and Learns: It reports on the outcomes of its actions and learns from the results to improve future decisions.

This continuous loop of requirement, measurement, decision, action, and reporting is what allows the network to adapt and evolve in real-time, without constant human intervention. The ultimate goal, as Niemöller explained, is to create a network composed of multiple "autonomous domains" – from the Radio Access Network (RAN) to the core network and cloud infrastructure – each managed by its own set of intelligent agents, all working in concert to fulfill the overarching business intent.

Part 2: Building Enterprise AI That Creates a Competitive Moat

The second presentation, delivered by Corentin Petit, shifted the focus from the theoretical architecture of autonomous systems to the practical challenges of building and deploying AI that delivers real business value. The core message was clear: off-the-shelf AI is a commodity; true competitive advantage comes from building a purpose-built AI that is deeply integrated with your unique data, expertise, and processes.

The AI Implementation Crisis

Petit began with a stark statistic: 85% of AI use cases fail to reach production. This isn't because the technology isn't powerful, but because enterprises face a number of significant hurdles:

Proprietary Data Not Integrated: Most companies' most valuable data is locked away in siloed, proprietary systems. Off-the-shelf AI models can't access or understand this data, limiting their usefulness.
Insufficient Accuracy and Performance: General-purpose models, while impressive, often lack the specific knowledge and nuance required for specialized business tasks.
Lack of Customization and Control: Enterprises need to be able to customize models to their specific needs and maintain control over their data and deployments.
Lack of Expertise: There is a significant talent gap in the industry, making it difficult for many companies to build and maintain their own AI systems.

This leads to a situation where many companies are stuck with "off-the-shelf AI" that is optimized for the wrong outcomes. It might be good at general conversation, but it can't perform the specific, high-value tasks that would truly transform the business.

The Path to ROI - Building Your AI Moat

The solution, according to Petit, is to build a "purpose-built AI" that acts as a "competitive moat" around your business. This involves a flywheel approach that transforms a company's unique assets into a long-term competitive edge:

Ingest: The process begins by ingesting your unique data, expertise, processes, and insights. This is the raw material that will make your AI unique.
Model: This raw material is used to train and customize a proprietary AI model. This isn't about building a foundation model from scratch, but about fine-tuning a powerful base model with your specific knowledge.
Deploy: The customized model is then deployed into your workflows and applications.
Capture: As users interact with the AI, you capture new data, feedback, and insights. This captured data then feeds back into the "ingest" phase, creating a virtuous cycle of continuous improvement.

The result of this flywheel is a long-term competitive edge characterized by:

Smarter Decisions: The AI can make decisions based on a deep understanding of your business.
Autonomous Operations: It can automate complex tasks, freeing up human employees for more strategic work.
Deeper Insights: It can uncover patterns and insights in your data that would be impossible for humans to find.
Continuous Improvement: The AI gets smarter and more valuable with every interaction.

To illustrate this concept, Petit presented a case study with CMA CGM, a global leader in logistics. The challenge was to automate parts of their complex order processing workflow. The solution involved a rapid, iterative process:

Discovery (2 weeks): The team worked with CMA CGM to understand their existing processes and identify the key pain points.
Build & Deploy (2 weeks): They quickly built and deployed an initial version of an AI agent to handle specific parts of the workflow.
Iterate & Evolve (4 weeks): They gathered feedback from users and continuously improved the agent's performance.

The result was a production application that was developed in just 8 weeks, leading to a 40% automation of the end-to-end workflow. This is a powerful demonstration of how a focused, purpose-built AI can deliver tangible results in a short amount of time.

Part 3: Building Effective Agents

The final presentation of this trilogy came from Christian Ryan of Anthropic, a company at the forefront of AI research and development. Ryan's talk provided a high-level overview of the evolution of agentic AI and offered practical advice for building effective and reliable agents.

The Agentic Evolution: From API Calls to Autonomous Agents

Ryan framed the current state of AI as the culmination of a three-stage evolution:

Simple API Calls: The first stage of applied AI involved using models for simple, discrete tasks like summarization, classification, and data extraction.
Workflows: The next stage involved chaining together multiple LLM calls and tools to accomplish more complex, multi-step tasks.
Agents: The current frontier is the development of autonomous agents that can take a high-level directive and then plan, act, and reflect to achieve that goal.

The key difference between a workflow and an agent is the "agentic loop." An agent doesn't just follow a predefined set of steps; it operates in a continuous cycle:

Plan: The agent observes its environment and the current context and formulates a plan.
Act: It takes an action in the environment using its available tools.
Reflect: It observes the outcome of its action, learns from the feedback, and adjusts its plan accordingly.

This loop is what gives agents their power and flexibility, allowing them to handle unexpected situations and learn from their mistakes.

Best Practices for Building Effective Agents

Building reliable AI agents is a significant engineering challenge. Ryan outlined several core principles that Anthropic follows:

Start Simple, Scale Up: Don't try to build a fully autonomous, do-everything agent from day one. Start with simple LLM calls for specific tasks, then build those into low-code workflows, and only then deploy agents when you need true flexibility and scalability.
Key Considerations: When designing an agent, it's crucial to think about the trade-offs between performance, latency, and cost. You also need to design for dynamic decision-making and, most importantly, "think like your agent." What information does it need to make the right decision at each step?
Design Principles: Maintain simplicity in your design. Prioritize transparency by showing the user what the agent is doing at each step. And finally, carefully craft the user interface to guide the agent and make it easy for users to interact with it.

The Importance of Rigorous Testing and Evaluation

One of the most critical and often overlooked aspects of building agents is testing and evaluation. Ryan highlighted several common mistakes:

Taking too long without empirical testing.
Overlooking data quality and infrastructure.
Relying on subjective "vibes-based" assessments rather than quantifiable metrics.

The best practice is to implement comprehensive evaluation frameworks, invest in telemetry to track the agent's performance, and design tests that are aligned with real-world use cases and validated with user-centric metrics.

Conclusion

The presentations from Ericsson, Mistral, and Anthropic, while different in their focus, all point to a shared vision of the future: a future where intelligent agents and autonomous systems are deeply embedded in our digital infrastructure and business processes.

From the macro level of self-managing telecommunication networks to the micro level of individual AI agents automating specific business tasks, the trend is clear: we are moving from a world where we tell our systems what to do to a world where we tell them what we want to achieve.

This shift represents a monumental opportunity, but it also comes with significant challenges. As we move forward, it will be crucial to embrace the principles of intent-driven design, purpose-built AI, and rigorous, iterative development to ensure that these powerful new technologies are not only effective but also reliable, trustworthy, and aligned with human values. The journey to true autonomy has just begun, and the insights from these leaders provide a valuable roadmap for the path ahead.