The AI agent market is experiencing a massive adoption gap that follows a power law distribution, and nowhere is this more obvious than the events of this past week. On one extreme, we had a heated technical debate between Anthropic and Cognition about single versus multi-agent architectures. On the other extreme, we got McKinsey’s embarrassing deck on AI agents that recommends models from 2022.
If you have been wondering what AI agents actually are and how to approach them strategically in 2025, this gap perfectly illustrates why most businesses are going to fail with their agent initiatives this year. The divide between those who understand the technical fundamentals and those falling for consulting buzzwords is enormous.
Let me break down what actually happened, why McKinsey’s recommendations are dangerous, and what strategic levers you need to focus on if you want your AI agent implementation to succeed.
What Are AI Agents Really?
Before diving into the drama, let’s establish the basics. An AI agent is simply an LLM plus tools plus policy guidance. That’s it. The agent can use tools, follow guidance, and operate with some degree of autonomy to complete tasks.
The current industry debate centers on whether you should build single-agent architectures or multi-agent systems. Single-agent advocates, like the team at Cognition who built Devon, argue that multi-agent systems introduce too much complexity and make it nearly impossible to maintain good production deployment standards.
Anthropic fires back with their Deep Research system, which uses multiple agents and claims vastly superior effectiveness. Their argument? Multi-agent systems are an efficient way to burn compute because they value correctness over simplicity.
The Token Burn Reality
Here’s where it gets technical, but stick with me because this matters for your strategy. If you do not burn enough tokens, you are unlikely to get to the correct solution with a large language model. The AI outputs tokens, and if it does not output enough of them, it often does not reach the solution.
This became embarrassingly clear in the recent Apple paper that claimed reasoning was dead. They did not allocate enough output tokens to make determining the solution computationally possible for at least one of their major puzzles. A large language model, Claude Opus, was even listed as a co-author on the paper that debunked Apple’s research over the weekend.
Anthropic’s multi-agent approach acknowledges this reality. They deliberately advertised a 7-hour runtime for Opus when they released it, because they are working on lengthening the horizon of agent-driven development and coding endeavors. They have the talent to implement multi-agent systems effectively, and they understand that burning computational tokens matters for solving complex problems.
Multi-agent systems burn more computational tokens but can achieve better correctness for complex problems.
McKinsey’s Embarrassing Recommendations
While Anthropic and Cognition are having sophisticated debates about computational architecture, McKinsey released a deck that’s frankly embarrassing. They are recommending models that are years old, like Claude Haiku. Nobody uses Haiku anymore. We use Sonnet and Opus. Why would you recommend two-generation-old models to CEOs who are presumably paying significant money for this advice?
They call state-of-the-art models things like Llama 3 8B, Gemini Nano, and Mistral Small. This is GPT-2 era thinking in what is almost a GPT-5 world. It looks like McKinsey’s technical teams either do not track actual model releases, or they think their clients are too uninformed to understand current capabilities.
Even worse, they are addicted to buzzwords. They talk about something called an “agentic AI mesh.” Ask any developer what an agentic AI mesh is. This is buzzword thinking with no technical substance. You can rebrand concepts all you want, but that does not make it innovation, and it certainly does not make it useful for decision-m.
McKinsey’s recommendations do not specify messaging protocols, state management schemas, or error handling patterns. These are fundamental technical considerations that determine whether your agent implementation succeeds or fails. Instead, they throw around technical buzzwords from 2022 and 2023, apparently hoping nobody will notice.
The Strategic Levers That Actually Matter
If you want your AI agent initiative to survive 2025, focus on these strategic levers instead of McKinsey’s nonsense:
Memory Architecture Design
Your memory architecture decision determines everything else. If you design memory and context access correctly from the start, you are doing the context engineering that enables you to shape instruction sets, policies, guidance, and the substrate of context that agents operate on.
This is where most implementations fail. Without proper memory architecture, your agents cannot maintain state effectively or learn from previous interactions. Context engineering is the key concept missing from most agent discussions.
Token Burn vs. Correctness Trade-offs
You need to understand the relationship between computational tokens and solution correctness. For complex problems, burning more tokens often leads to better outcomes. The question is whether the value of those correct solutions justifies the computational cost.
If you are working on problems where correctness matters more than cost efficiency, multi-agent systems might make sense. If you need fast, cheap solutions for simpler tasks, single-agent architectures could be better. But make this decision based on technical understanding, not consultant buzzwords.
Evaluation and Quality Measurement
This is the thing that separates successful implementations from failures. You can get everything else right – the build versus buy decision, the single versus multi-agent architecture, the statefulness, the memory design – but if you do not have proper evaluations, you are in trouble.
You need to measure model drift, understand how agents perform in production, and have systems in place to maintain quality over time. Without these measurements, your agent will degrade, and you will end up like the alarming number of companies reconsidering their agent investments right now.
The three strategic levers that determine agent implementation success.
The Power Law Distribution in Practice
The gap between the Anthropic/Cognition debate and McKinsey’s recommendations perfectly illustrates the power law distribution in AI agent adoption. On the far right side, you have teams with deep technical expertise arguing about computational architecture and token optimization. On the far left, you have consultants recommending outdated models and made-up concepts like “agentic AI mesh.”
Most businesses fall somewhere in the middle, which means they are susceptible to falling for consultant decks rather than understanding the fundamentals. This is why so many companies are reconsidering their agent investments. They have been told this would be easy, spent significant money, and are understandably frustrated with the results.
If this describes your situation, it is not your fault. But you can do better by understanding the actual technical considerations rather than relying on buzzword-heavy presentations.
Build vs. Buy Decisions
The hype around agents means there is money flowing into point solutions. If there are turnkey solutions like Zendesk integrations or tools like Lindy.ai that solve your specific problem, take them. You do not get extra credit for implementing a fancier agent – you just get more pain, and you better be sure the ROI justifies it.
Increasingly, there are off-the-shelf solutions that can handle common use cases without requiring you to build complex multi-agent systems or manage memory architectures yourself. The key is being honest about what you actually need versus what sounds impressive in meetings.
Looking at Current Tools and Platforms
When evaluating existing solutions, look for platforms that have solved the fundamental problems I have outlined. Some tools like the Minimax M1 Agent Platform are getting the research capabilities right, while others focus on specific verticals or use cases.
The key is understanding what problems these tools actually solve versus what their marketing claims. Most agent platforms are still maturing, and the gap between promise and reality can be significant.
Why This Matters for Your Business Strategy
The AI agent market is projected to be worth hundreds of billions in the coming years, which means the hype is going to continue. Let that hype work for you by being selective about what you build versus what you buy, but do not let it cloud your judgment about technical fundamentals.
An alarming number of companies are reconsidering agent investments right now, and it is largely because they fell for presentations like McKinsey’s rather than understanding the actual complexity involved. The companies that succeed will be those that either invest in real technical expertise or choose turnkey solutions that have already solved the hard problems.
The development teams arguing about single versus multi-agent architectures understand something important: this stuff is genuinely difficult. Memory architecture, state management, and evaluation systems are not buzzwords – they are technical realities that determine whether your investment pays off.
What To Do Instead
If you are looking at agents for your business, start with these practical steps:
First, understand your actual use case and whether it requires autonomous decision-making or if a workflow would work better. Many things companies call “agents” are actually just workflows, and workflows are generally better for most business processes.
Second, if you do need agents, focus on the strategic levers: memory architecture, token economics, and evaluation systems. These technical foundations matter more than whether you use the latest model or follow the latest consulting framework.
Third, be realistic about complexity. If Anthropic’s team with world-class talent finds multi-agent systems challenging to implement well, your team probably will too. There is no shame in choosing simpler solutions or off-the-shelf tools that work.
Finally, ignore consulting decks that rely on buzzwords and outdated model recommendations. The technical landscape moves quickly, and presentations recommending Claude Haiku in 2025 are not worth your time or money.
The AI agent space is moving rapidly, but success still comes down to understanding fundamentals and making informed technical decisions. Do not let the hype or consultant buzzwords distract you from what actually works.
The Role of Distributed Coordination in Multi-Agent Systems
For those leaning into multi-agent architectures, effective distributed coordination becomes paramount. This involves designing how multiple agents communicate, share information, and divide tasks to achieve a common goal. Without robust coordination mechanisms, multi-agent systems can become chaotic and inefficient, negating any potential benefits from increased token burn or correctness.
Consider a scenario where multiple agents are collaborating on a complex coding project. One agent might be responsible for understanding the high-level requirements, another for generating initial code snippets, a third for debugging, and a fourth for testing. If these agents do not coordinate effectively, they might duplicate efforts, create conflicting code, or miss critical errors. This is where well-defined messaging protocols and state management schemas become essential. McKinsey’s deck fails to address these practicalities, suggesting a fundamental lack of understanding of real-world agent deployment.
Interface Standardization and Protocol Differences
Another often overlooked aspect in the superficial discussions around AI agents is interface standardization. True, a CEO might not need to know the intricacies of every API, but they should grasp the importance of designing systems that can interact seamlessly. Different protocols underpin various AI models and tools. Ignoring these differences leads to integration nightmares and limits the scalability of agent solutions.
The debate between single-agent and multi-agent architectures also touches on this. A single agent stack might have a simpler internal interface, but a multi-agent system requires careful standardization of how agents exchange data and instructions. This is not just about technical elegance; it impacts the cost of maintenance, the speed of development, and the overall robustness of the system. Failing to consider interface standardization is a common pitfall for companies that rush into agent initiatives without a solid technical foundation.
The Trap of Vanity Metrics and Misguided ROI
Many businesses are caught in the trap of pursuing AI agent initiatives for the wrong reasons, often driven by a fear of missing out or a desire to appear cutting-edge. This leads to a focus on vanity metrics rather than true return on investment. If you cannot clearly define how an AI agent will reduce costs, increase revenue, or improve specific business outcomes, then your investment is at risk.
The frustration many companies are feeling with their agent investments stems from this misalignment. They were promised ease and transformative results, but without a clear understanding of the operational complexities and measurable KPIs, they find themselves with expensive, underperforming systems. This is why I emphasize evaluation so heavily. It is not enough to build an agent; you must be able to prove its value through rigorous measurement.
For instance, if an agent is designed to automate customer service inquiries, you need metrics on resolution rates, response times, customer satisfaction scores, and actual cost savings compared to human agents. Without these, it is impossible to determine if the agent is a success or a drain on resources. The lack of emphasis on such practical measurement in consultant decks is a glaring red flag.
The Importance of Talent and Expertise
The Anthropic team’s success with multi-agent systems is partly attributed to their exceptional talent. This highlights a critical, often understated, strategic lever: the quality of your technical team. Building and maintaining sophisticated AI agents requires deep expertise in machine learning, software engineering, and often domain-specific knowledge.
Companies that rely solely on external consultants or assume that off-the-shelf solutions will negate the need for internal talent are setting themselves up for failure. While turnkey solutions can be valuable for specific problems, true strategic advantage in AI agents comes from having the in-house capability to understand, adapt, and troubleshoot these complex systems. If you do not have the talent to manage token burn, design memory architectures, or set up robust evaluation frameworks, then even the best intentions will fall short.
The Future of Agentic Endeavor: Beyond the Hype
The current hype cycle around AI agents, while leading to some misguided investments, also points to a fundamental shift in how businesses will operate. AI agents, when implemented correctly, promise to automate complex, multi-step tasks that traditionally required significant human intervention. This can free up human capital for more creative, strategic, and interpersonal roles.
However, the journey to realizing this potential is fraught with technical challenges. The debates between single-agent simplicity and multi-agent complexity, the nuances of token economics, and the absolute necessity of robust evaluation systems are not academic exercises. They are practical considerations that determine the success or failure of real-world business initiatives.
My experience with intelligent automation frameworks has consistently shown that the true power of AI comes from a deep understanding of its underlying mechanisms and a meticulous approach to implementation. It is not about adopting the latest buzzword, but about applying foundational principles to solve concrete problems. This is why I stress the importance of understanding the fundamentals of agents: they are tools, they follow guidance, and they use LLMs. The complexity comes from how you combine and manage these elements.
In the coming years, we will see a shakeout in the AI agent market. Companies that built on shaky foundations, influenced by superficial advice, will likely abandon their efforts. Those that invested in genuine technical understanding and practical implementation will begin to see real returns. Your choice now determines which group you will be in.
