Fully Connected 2025 kickoff: The rise (and the challenges) of the agentic era

The evolution of AI is happening at an exponential rate, and understanding where we’ve come from is essential to navigating where we’re going. Quoting Shakespeare, “the past is prologue”—the events of the past few years inform the present reality and the future trajectory of artificial intelligence.

The Exponential Leap: From Past to Present

We are living through a period of unprecedented growth, with model sizes increasing 10x every 12 to 18 months. While this is an impressive input metric, the outputs are even more telling. Benchmarks are saturating faster than ever; older benchmarks took nearly five years to approach 100%, whereas newer ones like AI-M reached 80% in about a year.

This progress is not just academic; it’s manifesting in real-world applications and widespread adoption:

• Mainstream Adoption: Over 50% of new customers for platforms like Notion are using its AI features, indicating a shift from AI being an additive feature to an ambient part of the user experience. This trend extends to traditionally conservative fields like medicine, where a third of US physicians use platforms like Open Evidence for research summarization, and law, where over a third of top firms use Harvey for legal reasoning.
• The AI Coding Boom: The AI coding sector has exploded, generating over a billion dollars in annual run rate in the last year alone from companies like GitHub, Cursor, and Cognition.
• Dramatic Quality Improvements: The qualitative leap in generative AI is staggering. A comparison of AI-generated video from 2023 to 2025 shows a monumental improvement in coherence and realism. Similarly, music generation has evolved from basic tunes to complex, genre-specific compositions.
• A Generational Shift: AI is becoming a native tool for the younger generation. Examples include a five-year-old “vibe coding” a functional game and a university undergraduate systematically evaluating six different LLMs on five different tasks to optimize his studies in mathematics and economics.

The Modern Landscape: The Age of Agents, Reasoning, and Real-Time Learning

We are currently in an era defined by three key trends: reasoning, real-time capabilities, and rapid learning. A pivotal event was the release of DeepSeek-R1, which initially caused a massive single-day loss for NVIDIA’s market cap. The market misinterpreted this efficiency gain, fearing a reduced need for compute. However, this illustrates Jevons’s Paradox: as a technology becomes more efficient, its consumption increases because new applications become economically viable.

This has led to several key developments:

• Advanced Reasoning: As older benchmarks saturate, new, more challenging ones like Frontier Math are emerging to evaluate mathematical creativity and multi-step formal logic. We are on the verge of AI-driven scientific breakthroughs, with examples like the Korea Advanced Institute of Science and Technology using AI to discover new classes of hyper-elastic and thermoelectric materials.
• The Rise of Reinforcement Learning: The popularization of algorithms like GRPO from the DeepSeek model has led to an explosion in reinforcement learning. This enables the creation of smaller, custom, and highly efficient models tailored to specific domains.
• The Explosion of Agents: The ecosystem of AI agents has grown uncontrollably, with tools for every function and industry. The central question has shifted from “Can you build an agent?” to “Can you make it reliable, observable, and safe?”
  • Evaluation: New testing grounds like WebArena are being developed to evaluate agents on real-world tasks such as booking flights and debugging code.
  • Orchestration: The Model Context Protocol (MCP) is becoming the lingua franca for agent orchestration, allowing platforms like Weights & Biases to be turned into tools that agents can interact with to generate visualizations and reports.
  • A New UX Paradigm: We are witnessing a fundamental shift where “Agent Experience is the new User Experience.” APIs are becoming more important than UIs, as agents will be the dominant consumers of these interfaces. This is exemplified by the current conflict between Amazon, which wants users on its UI, and Perplexity, whose agent-based browser experience relies on API-like interactions.

Enduring Challenges in Production AI

Despite the rapid advancements, fundamental challenges—some new, some old—make productionizing AI incredibly difficult.

The New Challenge: Productionizing Agents

Agents are alluring and easy to demo but notoriously hard to productionize. The primary reasons for this are:

1. Radically Increased Expectations: User expectations have shifted from simple tasks (e.g., translating a word) to complex, nuanced operations (e.g., instantly translating a blurry, handwritten menu in a foreign country at midnight).
2. Unpredictability of Dynamic Systems: The real world is unpredictable, and there are countless ways for an agent to fail when operating in such an environment.

The Old Challenges That Remain

The core problems that existed years ago still persist and are, in many ways, amplified:

• AI Development is Not Software Engineering: The founding insight of Weights & Biases remains true. Software development is a linear process, whereas AI development is experimental and iterative. Most experiments fail, but the learning from those failures informs the next step.
• The Dizzying Pace of Innovation: With new models released almost daily, it’s a significant challenge to keep up, evaluate them, and determine which is best for a specific use case.
• The Need for a System of Record: In an industry with high job turnover, a company’s intellectual property is the accumulated learning from its experiments. Without a system to track this, knowledge walks out the door when a researcher leaves.
• The Iteration Loop is Still the Bottleneck: The ultimate goal is to accelerate the cycle of prototyping, iterating, deploying, fine-tuning, and redeploying. This loop remains the central bottleneck for AI teams everywhere.

• Tags:
• LLMs
• Reinforcement Learning
• Production AI
• Agents
• Model Evaluation
• AI Development