AI Agents Rising

Stochastic Disruption

We commented on this trend of 'intelligent stochastic execution' in prior posts. We've also explained the issues. But based on our most recent experience in building AI Agents, we believe we have arrived at a watershed moment with the technology.

AI agents aren’t just the next iteration of robotic process automation (RPA) as some claim—they’re a complete paradigm shift in how we build and think about software. Traditional RPA is great at automating repetitive tasks, but AI agents go further. They blend automation with cognition, turning software into something that reasons, plans, and executes with finesse. Instead of merely following instructions, we've built agents on LLMs which learn, adapt, and act independently. They don’t just do—they operate with finesse on top of logic engines.

This distinction is game-changing. Consider NVIDIA’s description of agents as systems that can reason through problems, create plans, and execute them using tools, or AWS’s framing of agents as self-determining programs that interact with their environments to achieve goals. Imagine replacing bloated, multimillion-line legacy software with compact AI agents that leverage LLMs as their operating systems. These agents use advanced probabilistic models like Markov decision processes (MDPs) and their more nuanced counterparts, partially observable MDPs (POMDPs), to make decisions in uncertain conditions. While early iterations faced accuracy issues due to the stochastic nature of LLMs, advancements in reinforcement learning and system-level safety evaluations are rapidly addressing these challenges, fueling the development of smarter, more reliable agents.

We anticipate a disruption in the legacy software market. Why code rigid systems when you can deploy a collection of nimble agents capable of evolving with new data and tasks? Open-source initiatives like Meta’s Llama 3.1 agentic system are already proving the concept. These agents break tasks into multi-step reasoning processes, integrate zero-shot tool learning, and dynamically adapt to new problems. In a market teeming with startups and tech giants alike, the rush to deploy agents is accelerating, and for good reason: they offer the promise of doing more with less—less code, less rigidity, and far less maintenance. The future of software isn’t just automation; it’s intelligence, and AI agents are leading the charge. As we posted recently, the seismic shift has started.

Now is the right time to get started with prototypes if you haven't already worked with the technology. What is important about Agent composition is identifying a technology stack and framework which is extensible and enduring, permitting your teams to address use cases with scale, while leaving room for experimentation -- and even rework. We're still at the front end of this disruptive trend, but it is a trend that will only gain momentum in 2025.