Agentic Era Part 2: How the Architectural Battle Between Model Maximalists and Code Craftsmen is Shaping AI's Future
Autonomous systems, not just bigger models, drive value creation in the era of autonomous AI. The debate over those systems' structures has big implications for businesses trying to build new moats.
Beneath the GenAI hype cycle, a fundamental transformation is unfolding.
We continue obsessing over prompts, tokens, benchmark scores, and the race to AGI. Meanwhile, the truly transformative shift that will create and destroy billions in market value is happening at the architectural level.
AI is evolving from tools we operate to workers that operate autonomously. This isn't just a product evolution; it's a fundamental rewiring of how technology creates value.
This architectural revolution is redefining:
How defensible moats are built when everyone has access to the same models
How SaaS businesses maintain margins when basic AI capabilities become commoditized
How traditional enterprises can extract asymmetric returns from AI investments
Which metrics matter when evaluating both AI-native and AI-enhanced companies
Last week, I kicked off a series analyzing this strategic inflection point and its implications for building durable businesses in what I'm calling the Agentic Era:
In this second installment, I examine the battle lines being drawn between competing visions for how these autonomous systems should be built. Today's architectural choices will determine whether companies capture value or get left behind as AI transforms from software into labor.
Introduction
OpenAI and Anthropic recently released comprehensive guides on building AI agents, each outlining distinct approaches to this emerging paradigm. Just a few days later, Harrison Chase, founder of LangChain, published a fierce rebuttal to OpenAI's perspective, igniting an industry-wide debate on the fundamental architecture of agentic systems. This technical confrontation unfolds against a backdrop of aggressive agent deployment by software giants, from Microsoft to Salesforce, and AI-native applications.
On Saturday, while walking past my building in Manhattan, I was struck by a giant advertisement that boldly proclaimed: "Stop hiring humans. Get AI sales BDRs instead!" from Artisan. In last month's piece on the "Next Productivity Frontier," I argued that AI would progressively replace human labor in cognitive tasks. That future appears to be arriving faster than anticipated.
Despite the criticism of first-case deployments and damaging missteps—consider Cursor's customer agent default last week (note: that has not been an obstacle to a massive fundraise at a 9B$ valuation announced yesterday!) —agents are nonetheless becoming an established part of the technology landscape.
This is the second article of a mini-series on the Agentic Era. In part one last week, I examined how orchestration, the backbone for agentic systems, has emerged as the critical competitive advantage in the LLM landscape.
Today, I analyze the next phase of this evolution: agentic systems that promise to release us from the prompt engineering treadmill and unlock unprecedented productivity gains by transforming AI from software we operate into workers that operate autonomously.
What are agents? Beyond prompts to purpose
AI agents are systems designed to understand requests, reason through problems, and take actions independently to accomplish goals. They range from simple chatbots to complex systems that can research topics, make reservations, or manage digital tasks.
There is no consistent definition of an agent in the industry. The technology giants offer different perspectives on what constitutes an agent, as evidenced by their recently released guides.
For OpenAI, "Agents are systems that independently accomplish tasks on your behalf."
For Anthropic: "Agents can be defined in several ways. Some customers define agents as fully autonomous systems that operate independently over extended periods, using various tools to accomplish complex tasks. Others use the term to describe more prescriptive implementations that follow predefined workflows. At Anthropic, we categorize all these variations as agentic systems but draw an important architectural distinction between workflows and agents: Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."
The latter shows that agentic systems are a combination of agents and workflows.
An important caveat deserves attention: agents are not always needed. At least for now. So not all software has to be agentic.
Per Anthropic: "When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all. Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense."
What matters most in this evolving landscape is that we're witnessing the birth of AI systems that can initiate and execute complex tasks with minimal human oversight. The prompt-response paradigm is evolving into goal-directed behavior, from reactive to proactive AI. The value creation occurs at the intersection of these components, not in either extreme.
The technical battleground: Model maximalists vs. code craftsmen
A heated philosophical battle is playing out that will determine the future trajectory of AI investments. This great agent debate has split the industry into two opposing camps with radically different visions.
On one side, the "Model Maximalists" place their faith in the raw power of ever-larger foundation models. This "Big Model Take the Wheel" philosophy, evangelized by Applied AI experts like Rahul Sengottuvelu at Ramp, argues that as models scale, they internalize the structure we currently build externally. In a video that went viral, he advocates for the "LLM takes it all approach."
This is essential viewing for understanding the debate:
As O-team researcher Hyung Won Chung explains, " What we see over and over is that the method with the more structure typically starts off quickly because the structure itself serves as a shortcut, and that serves us well until it doesn't. And it doesn't really scale up. In contrast, the one with less structure typically starts off and doesn't work because we give too much freedom to the model and it doesn't know how to utilize it until we give enough data and compute and good algorithm to leverage that. And at some point, it just gets better and better, and we call this a more scalable solution."
The allure of this approach is seductive: minimal engineering effort, maximum flexibility. But the economics tell a different story, and Sengottuvelu acknowledges this limitation. Current implementations demand up to 10,000x more compute than structured alternatives—a cost equation that breaks in all but the most valuable use cases. At a time when compute remains restricted—as evidenced very clearly by Microsoft and Meta's earnings calls last week—and cloud providers are reporting margin compression from AI workloads, this matters tremendously.
The opposing "Code Craftsmen," led by Chase, advocate for "writing more code" to create structured frameworks that guide agent behavior. This approach replaces general-purpose LLM calls with purpose-built agent components, delivering predictability, explainability, and cost efficiency—the trifecta that enterprise buyers demand.
This debate transcends academic posturing and is at the core of building viable agentic systems—agents that actually work and produce value. Horizontal applications like coding assistants may benefit from the Model Maximalist approach, but vertical solutions in finance, healthcare, and manufacturing demand a structure that encodes domain-specific logic and compliance requirements.
The Context Moat: The True Value Creator
As the industry debates the merits of model-centric versus workflow-centric approaches, a more fundamental source of competitive advantage is quietly emerging: the context moat.
The context moat refers to the durable advantage created by applications that own and maintain the contextual layer of AI interactions – the accumulated knowledge, preferences, history, and state that persists across user sessions and model improvements.
This is distinct from merely having access to frontier models (which trends toward commoditization) or building elegant workflows (which face obsolescence with model improvements). The context moat is about owning the layer where meaning and continuity reside.
What makes the context moat so powerful is that it addresses the fundamental limitation of large language models: their statelessness. Without persistent context, even the most advanced models start each interaction from scratch, unable to build meaningful continuity or truly understand a user's evolving needs.
Applications that effectively capture, maintain, and leverage context create compounding advantages that become increasingly difficult to replicate over time. This context layer isn't just data – it's the accumulated understanding of users, tasks, and domains that grows more valuable with each interaction.
Consider Notion AI, which has built a powerful moat through its integration with users' knowledge bases. By leveraging the context of documents, projects, and historical usage patterns, it creates an agent experience that becomes increasingly valuable over time. Each interaction enriches the context, making the system progressively more aligned with user needs.
Similarly, Bloomberg has extended its terminal ecosystem with agentic capabilities that utilize the firm's vast financial datasets and proprietary analytics workflows. This combination of rich context and specialized workflows creates a compelling value proposition that would be impossible to replicate without similar-scale data assets.
The Orchestration Framework Landscape: Tools Shaping the Agentic Future
As the agent paradigm matures, a dynamic ecosystem of orchestration frameworks has emerged, each offering distinct capabilities for different use cases. Building on Harrison Chase's comprehensive analysis (full comparison here), we examine three leading solutions that highlight the diverse approaches to agent architecture:
LangChain's LangGraph: A graph-based orchestration layer ideal for complex workflows. Its cyclic execution and observability tools (via LangSmith) enable enterprises like Klarna to manage customer support for 85 million users, reducing resolution time by 80%. Its complexity, however, can increase development time.
OpenAI's Agents SDK: A lightweight framework prioritizing simplicity, with robust guardrails and tracing for enterprise needs. It's well-suited for rapid prototyping (e.g., building chatbots) but may lack flexibility for highly customized workflows.
LlamaIndex: Evolving from a data framework, it excels in knowledge-intensive applications, such as document search for legal firms, by integrating large knowledge bases. Its strength in data-heavy tasks comes at the cost of higher latency.
Another notable approach comes from CrewAI, which focuses on multi-agent coordination by creating "crews" of specialized agents that collaborate on complex tasks. This framework is particularly promising for workflows that require diverse expertise and parallel processing capabilities.
The diversity of these frameworks reflects the early stage of the market, with various architectural patterns competing for dominance. Organizations evaluating these frameworks should prioritize flexibility, observability, and alignment with their technical capabilities and deployment requirements.
The current state of agent frameworks indicates they're increasingly converging on the importance of the context layer. Looking at the Agent Frameworks above, we can see that Memory – a key aspect of context management – is central to effective agent design.
The transformation ahead
Drawing on observations from enterprise AI deployments, several key developments will shape the agentic landscape in the coming months:
First, the Model Context Protocol (MCP) introduced by Anthropic could play an important role as a standardized interface for agentic systems. This open protocol—which I plan to explore in the next installment of this series—aims to provide a consistent way for AI applications to connect with data sources and tools. Rather than claiming to have all the answers, MCP represents one thoughtful approach to addressing the integration challenges that currently limit agent capabilities.
Second, economic pressures will likely drive pragmatic convergence between the two technical camps I've described. Capital constraints and margin considerations suggest that hybrid architectures—using foundation models for reasoning while implementing structured workflows for execution—may find the most practical balance between capability and efficiency. Hence, reinforcing the need for agentic frameworks.
Last, organizations that develop effective memory systems to preserve contextual understanding are building meaningful competitive advantages. An agent that understands your business processes becomes increasingly valuable over time as it accumulates institutional knowledge, potentially creating significant switching costs. These memory systems represent a critical, often overlooked component of agentic architecture that will determine long-term success.
For investors and executives navigating this emerging landscape, the implications deserve careful consideration. The Model Maximalist approach may reduce R&D expenses, but creates operational cost structures that don't scale efficiently. Structured frameworks require more upfront investment but deliver better economics at scale—an important factor as AI moves from experimentation to production deployment.
Success in the agentic era will hinge on integrating AI into business processes, with orchestration as the linchpin for coordinating agents, workflows, and tools. Executives should prioritize agentic frameworks that align with their domain's needs and invest in memory systems to build contextual moats. Investors should focus on companies leveraging proprietary data and workflows for defensible growth.
Start small by experimenting with agentic pilots in high-value workflows. Scale strategically to capture the productivity gains of this transformative era.