Deep|LLM 1Q25:AI Agents Are Going To Transform The Entire Investment Landscape

From Copilots to Agents: The Dawn of Autonomous AI Reshaping Global Markets

This comprehensive report provides an in-depth analysis of the rapidly evolving frontier AI landscape, with a particular focus on the emergence of AI agents and their transformative impact on the global investment ecosystem. Drawing on extensive conversations with frontline LLM researchers, we present a uniquely solid and detailed examination of the current state and future trajectory of artificial intelligence.

The report explores the significant shift in AI model development—moving from general capability improvements to practical, task-specific applications with real-world impact. We analyze how AI agents, particularly in coding, are creating unprecedented commercial opportunities and reshaping entire industries. Our research reveals that the field is entering what we call the "second half" of AI development, where success is increasingly defined by task design, system architecture, and practical utility rather than generalized intelligence.

Through rigorous analysis and privileged insights from key industry participants, we provide investors with critical perspectives on how coding agents are emerging as the first viable AGI use case with substantial market potential. The report examines the technical innovations driving this transformation, including advances in long-term memory, tool utilization, planning, and autonomous execution capabilities.

Our findings suggest that 2025 will be a breakout year for AI agents, with leading models like Claude 4 and the upcoming GPT-5 driving a paradigm shift from assistive "Copilots" to autonomous "Companions" capable of end-to-end task completion. This transition carries profound implications for productivity, labor markets, and investment opportunities across the global economy.

This report is essential reading for investors, technologists, and business leaders seeking to understand and capitalize on what may be the most significant technological and economic transformation of our generation.

Table of Contents

• Trends in AI Model Development

• Evolving Perspectives on Model Capabilities and Compute Demand

• AI's Second Half: Coding Leads the Way in Real-World Deployment

• The Next Breakthrough: Learning by Doing (Online Learning)

• Systematic Development of Agent Architectures

• Commercial Potential and Deployment Outlook

• Assessment of Model Company Progress

• Frontier AI Model Updates

• Strategic Positioning by Major Players

• Anthropic's Contribution to AWS Growth

• Meta's AGI Strategy

• The Rise of AI Agents: 2025 as a Breakout Year

• The Rise of Coding Agents

• OpenAI Codex and Claude 4 Analysis

• The Strategic Impact of Coding Agents

• Google's AI Strategy and Compute Demands

• Multimodality AI and World Model Development

Trends in AI Model Development

• AI model development is entering a new phase. The focus has shifted from initial concerns about models' limited capabilities—largely dictated by pretraining scaling laws with data constrain—to an emphasis on reinforcement learning and test-time scaling techniques. Now, the conversation is evolving again, reflecting on the inherent limits of pre-training and moving towards a more task- and application-driven "second half" for AI.

• The current trend signifies a key transition: from pursuing general-purpose performance improvements to achieving tangible real-world impact in specific, high-value domains—most notably through the emergence of coding agents.

• Coding stands out as a natural area for a breakthrough. Its clearly defined tasks, robust feedback loops, and self-reinforcing data flywheel make it a prime candidate for the practical deployment of large language models.

• There's also a growing push to systematically enhance agent capabilities—including long-term memory, tool utilization, planning, and execution. AI systems are increasingly evolving into agents that learn by doing.

• Model iteration is accelerating at a breakneck pace. Industry leaders such as Google, Anthropic, and OpenAI are now delivering major upgrades every 3–6 months, fueling intense competition.

• Commercial prospects are rapidly coming into sharper focus. While agents haven't yet had their "ChatGPT moment," coding agents are already showing promise in automating certain aspects of human work—hinting at enormous commercial potential.

Evolving Perspectives on Model Capabilities and Compute Demand in AI

• In early 2024, the prevailing view was that AI applications were mainly held back by the limited abilities of foundation models. The dominant belief was that, simply scaling up—guided by the Scaling Law—was the key to unlocking progress. Market sentiment was incredibly optimistic; Nvidia's stock surge was widely interpreted as confidence that bigger models and more computing power would directly lead to broader and more capable applications.

• By July 2024, the cracks in the Scaling Law narrative started to show with the exhaustion of pretraining data. Then the community's attention pivoted toward reinforcement learning, particularly the potential of post-training to improve model performance.

• Around December 2024, a few months after o1's release, a more skeptical and cautious tone began to emerge. Post-training was yielding only marginal gains—especially when it came to generalization. Real progress remained largely confined to areas like coding and math, and synthetic data for pretraining failed to live up to the hype, adding to the growing unease.

• These limitations sparked a wave of critical reflection on the pre-training paradigm. Following DeepSeek's release in Jan. 2025, the debate centered on whether pre-training truly required such massive computational resources, and whether reinforcement learning or self-play could be meaningfully extended beyond coding and math. Back then, there was growing optimism that self-play and RL might offer a more scalable path forward.

• However, today, we believe the field has reached a new inflection point. The next wave of progress may not depend on fully generalized intelligence. In fact, reinforcement learning doesn’t necessarily require broad generalization. In vertical, well-bounded use cases like coding agents, large models are already delivering substantial commercial value. These settings resemble Tesla’s FSD system, where the data flywheel drives continuous improvement. Through practical applications—coding, agents, and computer use—AI models are achieving meaningful real-world deployment. This marks a transition into the "second half" of AI.

• When the conversation revolves around pre-training breakthroughs, AI is still in its "first half," fixated on improving the model itself. But when the focus shifts to agent-based applications, it signals entry into the "second half," where success hinges on task design, system architecture, and practical utility.

• The past was about general-purpose AI and improving generalization. Now, it's about focusing on high-value tasks that AI can do consistently, clearly defining the scope of those tasks, and creating meaningful ways to evaluate performance.

AI’s Second Half: Coding Leads the Way in Real-World Deployment

The current wave of AI innovation is centered on building systems that can learn continuously, remember persistently, and improve autonomously. Core technical areas include seamless human-AI interaction, long-term memory architectures, and new paradigms for task definition and evaluation. At the center of agent development lies one pivotal goal: building an AI “brain” that evolves with use.

Coding stands out as the ideal proving ground for this vision, offering several advantages:

• It inherently involves human-computer interaction.

• Each program execution provides continuous feedback.

• There's a constantly evolving codebase, facilitating long-term memory accumulation.

• Task definitions and evaluation criteria are extremely clear.

Therefore, coding has become the first successful real-world application of the "second half" of AI. More importantly, coding possesses a data flywheel mechanism similar to that of self-driving cars: when programmers use AI tools for coding, they are, in effect, feeding back into the model, providing valuable feedback data. In contrast, while chatbot scenarios can collect user chat data, this data often cannot be directly used for training and may even lead to a decline in model capabilities.

The Next Breakthrough in AI Models: Learning by Doing, a.k.a. online learning

A major challenge ahead for large-scale AI systems is developing the ability to learn by doing—updating their parameters in real time as they infer, reason, and act. This mirrors human on-the-job learning and could unlock unprecedented levels of adaptability and autonomy.

Online Learning is already widely used in recommendation systems. Models can adjust recommendations in real-time based on immediate user behavior. Whether this mechanism can be replicated in LLMs will be a critical area of research.

Specifically, in the AI field, unlike recommendation systems that only update user vectors, LLMs also need to update their overall weights. Therefore, how to incorporate Online Learning into the training of LLMs and Agents is a key research topic for Frontier Research Lab.

Systematic Development of Agent Architectures

While some effort still goes into pre-training, the majority of resources are now concentrated on deploying agents, and the field has already seen substantial progress.

Even though challenges like self-play and generalization remain unresolved, real commercial value is emerging in applied domains such as coding, software engineering, and agent systems. Model development is increasingly focused on building robust agent architectures, featuring:

• Memory Module: Long Context technologies are used to mitigate short-term memory limitations, and the Web provides long-term memory support.

• Tool use Capabilities: The MCP module manages the invocation processes of various tools, allowing the model to automatically write code to call APIs, thereby enhancing its execution capabilities.

• Multimodal Perception: World model architectures and similar approaches enhance the model's understanding of the external world and code.

• Execution and Reasoning Capabilities (Instruction Following, Planning): Recent advances in models like Claude and Gemini 2.5 have pushed the boundaries of instruction-following, planning, and reasoning capabilities.

Ultimately, nearly every foundational breakthrough in model training over the past year serves one purpose: solving the structural challenges of real-world AI agent deployment.

Commercial Potential and Deployment Outlook