Introduction
Agentic AI has moved at remarkable speed. In just a year, what started as proof-of-concept demos and experimental pilots has turned into real deployments across industries. The idea of autonomous systems that can reason, act, and execute multi-step processes is no longer science fiction, it’s becoming part of enterprise operations.
But as many organizations have discovered, building functional agentic systems is far harder than talking about them. The hype often obscures the fact that most pilots struggle to move beyond the demo stage. The gap between what looks impressive on paper and what delivers value in production is where the real work lies.
Recently, McKinsey published an article summarizing six lessons from their first year of hands-on work with agentic AI. Their insights capture much of what we’ve seen as well: the importance of focusing on workflows, building trust through evaluation, and recognizing that humans remain central in the loop.
At Predli, we’ve spent the past year developing agentic AI for enterprises across different domains. Some of our discoveries mirror McKinsey’s, but we’ve also learned additional lessons.The kind that only emerges when you’re solving real problems under real constraints. Below, we revisit McKinsey’s six lessons and expand with our own observations.
Predli’s Lessons from the Field
1. Framework Choice is Strategic
The agentic ecosystem is evolving fast - LangChain, LangGraph, CrewAI, Autogen, OpenAI’s frameworks. Each promises power, but each also brings complexity. We’ve seen organizations adopt popular frameworks only to discover they don’t align with the actual problem, or worse, that they’ve locked themselves into a structure they can’t easily adapt. The right framework is not the most hyped or feature-rich. It’s the one that fits the use case, matches the team’s capacity to maintain it, and leaves room for adaptation as the ecosystem matures.
2. Keep Architectures Lean
Agent architectures range from simple ReAct models to deep multi-agent orchestration systems. The temptation is always to add more layers, more complexity, more “intelligence”. But more is not always better. We’ve repeatedly seen leaner architectures outperform more elaborate ones, both in terms of performance and maintainability. A simple setup is easier to debug, cheaper to run, and faster to iterate. Complexity should be a response to a real need, not the default starting point.
3. Evaluation is the Backbone of Trust
There’s no universal metric for evaluating agentic systems. For one workflow, it might be accuracy or retrieval precision. For another, it might be user trust, reduction of hallucinations, or domain-specific outcomes. The key is to embed evaluation into the system itself, not bolt it on afterwards. Every iteration, whether it’s a prompt change, an architectural tweak, or a new integration, should flow through an evaluation pipeline. Without this, teams can’t separate genuine progress from regression.
4. Design for Scalability and Generalization
Specialized agents solve immediate problems but often become brittle. When workflows evolve, they break. By contrast, systems designed with generalizability in mind adapt to new contexts and extend beyond their initial scope. The long-term value lies in agents that can scale across adjacent use cases. In practice, this means structuring agents in a way that allows them to be reused, extended, or repurposed, without rebuilding from scratch.
5. Prompting is a Craft, Not a Side Note
Prompts look simple, but they are often the hardest and most time-consuming part of building agents. Poor prompting creates brittleness and unexpected failures. At Predli, we’ve learned that prompts need to be designed, stress-tested, and iterated systematically. Prompting is not wordsmithing, it’s architecture. It’s how you align the agent’s reasoning with the workflow’s demands.
6. The Community is Writing the Playbook
Agentic AI is too young for any single company to have all the answers. Many of the practices we now consider standard, step-by-step reasoning, external memory, human-in-the-loop workflows, came from open experimentation across the community. The pace of change is so fast that no one can afford to innovate in isolation. Learning from shared successes and failures is not optional, it’s how progress happens.
McKinsey’s Six Lessons - and How They Connect
When McKinsey shared their six lessons from agentic AI deployments, we found much to agree with. Their findings reinforce several of the themes we’ve just described, while adding emphasis in other areas.
They remind us that it’s not about the agent but about the workflow: something we’ve also seen repeatedly in our projects, where embedding agents into redesigned processes makes the difference between a flashy demo and genuine adoption. They note that agents aren’t always the answer, and that in some contexts traditional automation is still the smarter choice.
Their strong focus on evaluation resonates deeply with our experience: without rigorous evaluation, trust evaporates quickly. Similarly, their call for observability at every step connects closely with our belief in transparent, lean architectures.
McKinsey also highlights reuse as a key driver of scalability, which aligns with our principle of designing for generalization rather than one-off solutions. And finally, they emphasize that humans remain essential. Agents can handle speed and scale, but people are still needed for oversight, context, and decision-making.
Taken together, McKinsey’s lessons and our own form a coherent picture: the most effective organizations are those that balance ambition with discipline, building systems that are both trustworthy and adaptable.
Where This Leaves Us
One year in, a pattern is clear. Agentic AI is moving beyond experimentation and into real business operations. But the organizations that succeed are not those chasing the most complex architectures or the flashiest demos. They are the ones approaching agentic AI with pragmatism, iteration, and human-centered design.
What excites us is how quickly the field is maturing, not because it’s eliminating challenges, but because the community is learning faster together than any single team could alone. The next year will be about scaling responsibly, strengthening governance and trust, and finding the balance between human creativity and agentic automation.
Agentic AI is no longer just an experiment. It’s becoming an asset, and those who design with discipline today will be the ones defining how it creates value tomorrow.
