The 'Year of Agents': Reinforcement Learning, Agent Collaboration, and the Future of Data Analytics

Biweekly Data & Analytics Digest: Cliffside Chronicle

Author

Josh Miramant
29 Jan • Estimated Reading Time: 6 minutes

Deepseek R1 enters the chat: Betting (and winning) on Reinforcement Learning

DeepSeek R1, a new reasoning model that claims to rival OpenAI’s o1 across math, coding, and general reasoning tasks. It’s open source, far cheaper, and even lets you see its chain of thought. But I believe the real game-changer isn’t any of these features. It’s what they represent: the power of a pure reinforcement learning (RL) approach.

DeepSeek R1 is a glimpse into how an LLM can learn to “think” and refine its reasoning skills without needing an enormous dataset of Chain-of-Thought examples. It reminds me of AlphaGo’s story. Back in 2016, DeepMind showed how pure RL could beat the world’s top Go player by relentlessly training against itself. Now, DeepSeek is applying that same principle to language models. Whether you’re following along for the open-source angle or the 90–95% cost savings, the real takeaway is that RL-driven LLMs may just be the future of AI.

Google: Chain-of-Agents outperform RAG and In-context

Google’s new “chain-of-agents” approach, which is already outperforming both Retrieval-Augmented Generation (RAG) and in-context solutions for long-context tasks.

In a nutshell, the chain-of-agents method taps into multiple specialized LLMs working together as agents, each taking on a piece of the puzzle, passing the output forward, and iterating until they converge on a strong solution. This collaborative approach has been especially effective for tasks where a single model might struggle to keep track of extensive or evolving information. Learning from the paper:

  1. Collaboration beats isolation: Letting models “divide and conquer” can yield better outcomes than relying on a single LLM prompt or a static knowledge retrieval approach.

  2. Long context management: When data is spread across a large corpus or a conversation that evolves over time, chain-of-agents helps ensure no critical information gets lost.

  3. Scalability and adaptability: Each agent can be fine-tuned for different types of tasks, so the whole system can seamlessly adapt to various problem domains.

This is a big step toward more powerful, flexible LLM ecosystems. Instead of having one model juggle everything, we’re moving closer to teams of AI agents collaborating in real-time.

OpenAI release Operator. A ‘Computer-Using Agent’…and it kinda works.

OpenAI Operator is a tool that enables seamless integration of OpenAI’s models into cloud-based VM with access to web browser…it’s demoed showing it booking flights (not great), booking reservations (okay) and doing other custom research.

On the impressive side, it handles multi-step instructions with surprising ease and shows real potential for freeing us from routine tasks. However, it also revealed rough edges. The agent occasionally became stuck on simpler commands, and figuring out how best to chain tasks together took some trial and error. Even so, I’m convinced these early limitations will be ironed out in future versions. As OpenAI refines how the AI decides between different actions, we’ll likely see increasingly reliable operators capable of handling more complex work.

This is definitely worth playing with…in future iterations.

Structuring the unstructured…data

With more and more enterprise data is made up of text, audio, and images that don’t fit neatly into rows and columns. In fact, IDC predicts that by 2025 the global datasphere will reach 175 zettabytes and around 80% of that will be unstructured. That sheer volume is already driving new AI tools and methods focused on search, classification, and advanced analytics that can handle messy or free-form information.

As unstructured data continues its explosive growth, there’s a massive opportunity to derive insights and competitive advantages. Traditional BI tools have always struggled here, but AI is bridging that gap through breakthroughs in language models, image recognition, and more. It’s also creating new challenges around data management, security, and governance, which the article notes will require careful planning.

Whether it’s customer support transcripts, social media posts, or sensor logs, understanding and acting on that data is quickly becoming table stakes for modern organizations.

Shameless plug: this is an area we have a lot of experience with at Blue Orange.

Blue Orange Blog Spotlight: The Role of AI in Data Governance

As data governance becomes essential to evolving business needs, organizations are searching for ways to manage it more efficiently, securely, and ethically. Data governance has become more than a compliance checkbox—it’s a strategic imperative.

What topics interest you most in AI & Data?

We’d love your input to help us better understand your needs and prioritize the topics that matter most to you in future newsletters.

Login or Subscribe to participate in polls.

“You can have data without information, but you cannot have information without data.”

– Daniel Keys Moran