Mike’s Agent Insights #6

Welcome to the 6th edition of Mike’s Agent Insights coming to you from Alex’s wedding in Colorado.

I have appreciated all of the conversations which the last couple of posts have started,  I am looking forward to sharing even more. 

In a world where AI increasingly shapes our lives and AI agents will increasingly transform even more aspects of our day to day  lives, it's important to stay informed and understand the implications of this technology. That's why I'm launching Mike's Agent Insights, a newsletter dedicated to exploring the latest research, investment, advancements and applications of AI agents. 

This week I want to explore; 

1) New agent products in the news

2) Agents research

3) Thought pieces on agents

4) Honorable Mention, Open Source Breakthroughs

New agent Products in the News

Perplexity's AI-powered shopping assistant simplifies product research and purchases - (Perplexity Pro Shop)

The new features empower users with faster, easier, and more reliable online shopping, saving time while offering unbiased, AI-driven product recommendations. Perplexity combines AI technology with integrations like Shopify to create a seamless shopping experience. Features include one-click checkout for U.S. Pro users, visual search tools, and unbiased product comparisons in clear language.  

Image Source: Reddit

This innovation marks a shift toward AI-driven commerce, enhancing convenience for shoppers while offering merchants new tools to connect with consumers and leverage data insights.  

• One-click checkout: U.S. Perplexity Pro users can save time with integrated purchasing and free shipping.  

• Snap to Shop: Visual search identifies products from photos for quick discovery.  

• Unbiased recommendations: AI-powered product cards provide clear comparisons, free from sponsorship.  

• Merchant benefits: Free API access, payment integration, and a dashboard for sales insights.  

• Future expansion: Global rollout and additional features are planned to scale the shopping experience.  

My thoughts: This one is really exciting to me.  The product itself is unique to what has been released to the market so far. The other thing is that they are bringing in people that have shops and e-commerce sites to be able to participate as well. The screenshots I could find about the product are interesting and I am really wondering what the discovery part of the product looks like. How do I get to the place where I can actually purchase? Perplexity has been a standout in search, so it is interesting to see them stepping into commerce instead of ads which has been the real revenue driver for search.  This could be the first step in a big transformation in search. 

Microsoft introduces Copilot Actions, new agents, and tools to transform productivity and IT management - (Spataro @ Microsoft)

Microsoft’s updates to Copilot empower employees with personalized AI assistants, streamline business processes with new agents, and equip IT teams with advanced controls for secure and efficient AI adoption. At Microsoft Ignite 2024, new features like Copilot Actions and advanced agents were unveiled, enhancing Microsoft 365’s ability to automate tasks, access organizational knowledge, and provide real-time language interpretation in Teams meetings.

These innovations position Copilot as a central tool for scaling productivity, modernizing workflows, and enabling IT teams to lead AI-driven transformation confidently. Watch the demo video. 

• Copilot Actions: Automates repetitive tasks with simple prompts, like daily action summaries and meeting prep.

• New agents: Includes SharePoint agents for quick document searches, Interpreter agents for real-time language translation, and Employee Self-Service agents for HR and IT tasks.

• IT management: Copilot Control System provides tools for data protection, user governance, and measuring AI ROI with Copilot Analytics.

• Improved tools: Teams Copilot summarizes visual content, Outlook Copilot schedules meetings, and PowerPoint Copilot offers multi-language translation.

• Expanding ecosystem: Partners like ServiceNow and Workday are adding agents to enhance functionality across HR, finance, and sales.

My thoughts: The demos videos show some big shifts in paradigms.  Making agents shareable across MSFT products and making them collaborative.  I also think the governance is a standout from the other early entrants in this space.  Competitively, this is worlds ahead of where Google is with their current Gemini integrations in GSuite products. I 

Sema4.ai empowers businesses to deploy AI agents in minutes, transforming workflows - (Plumb @ VentureBeat)

Sema4.ai’s no-code platform enables non-technical business users to create and manage AI agents, automating complex tasks and driving efficiency without relying on IT or developers. Founded less than a year ago, Sema4.ai has rapidly emerged as a leader in enterprise AI. With $30.5M in funding, it has piloted its platform with six Fortune 2000 companies, demonstrating significant productivity gains in areas like invoice reconciliation and compliance. 

AI agents are reshaping enterprise operations, with platforms like Sema4.ai enabling a shift from IT-driven processes to business-user-led automation. This shift enhances scalability, flexibility, and innovation across industries.

• No-code design: Features tools like natural language runbooks and document intelligence for easy adoption.

• Interoperability: Works with leading LLMs, including OpenAI, Claude, and Azure, for maximum flexibility.

• Proven ROI: In early trials, agents performed over 80% of knowledge work tasks autonomously.

• Real-world success: Koch Industries automates invoice reconciliation, saving time and resources.

• Future potential: Applications range from market analysis to external communications, promising scalable impact.

My thoughts:  Starting to see a bunch of these in investor presentations and new funding announcements.  Sema4.ai caught my eye because of having a few real world stories already but also they're interruptibility pitch.   Coming up one of these weeks I'll try and do a side-by-side of a bunch of these similar no-code business automation agent platforms.

Agents research

Magentic-One: A powerful multi-agent system tackling complex, multi-step tasks - (Fourney, Bansal et al @ Microsoft)  

Magentic-One advances AI by enabling agents to autonomously handle intricate workflows across domains, from coding to web navigation, while addressing risks tied to agentic systems. Magentic-One, developed by Microsoft researchers, leverages a multi-agent architecture where an Orchestrator manages specialized agents like a Coder and WebSurfer. It achieves state-of-the-art performance on benchmarks like GAIA and AssistantBench, pushing the boundaries of generalist AI.  

Image Source: Magnetic One Github

As AI evolves from generative models to agentic systems, tools like Magentic-One highlight both opportunities and challenges. These systems promise to revolutionize productivity but require robust safety and governance measures to mitigate risks.  

• Agent architecture: Includes an Orchestrator and agents for coding, browsing, file handling, and executing tasks.  

• Performance: Matches state-of-the-art benchmarks, showcasing adaptability without modifying core capabilities.  

• Open-source tools: Built on Microsoft’s AutoGen framework, with AutoGenBench for rigorous evaluation.  

• Risks: Real-world tests revealed potential for undesired actions, emphasizing the need for human oversight.  

• Future focus: Research on reversibility of actions and phishing/misinformation defenses is critical to safer agentic AI.  

My thoughts:  Microsoft research has been on a roll recently and this is a good follow on paper to some of the other multi-agent research projects they have published recently. Personally, I think this is very representative of how we should expect cloud providers to provide access to a series of agentic base tools to solve a wide variety of problems for developers and enterprises.   I also think this model allows for a blend of capabilities from different providers and services.  I think the industry is at some point going to outgrow getting everything from one closed source provider and start looking for hybrid approaches, especially in IT where that shift from monolithic architectures in the past is something that gradually got torn down as people invested in the things that made the most sense for their business to solve for different pieces of their overall IT architecture.

Thought pieces on agents

Architecting data stacks for AI agents is key to unlocking enterprise innovation. - (Sharma @ Venture Beat)

AI agents are transforming enterprise workflows by automating complex tasks, but their success hinges on robust, well-designed data strategies that ensure accuracy, adaptability, and governance. Since the launch of ChatGPT in 2022, generative AI has revolutionized business applications. The current wave of AI agents is evolving beyond prompt-based systems into autonomous entities capable of executing tasks and making decisions, supported by advances in large language models (LLMs) and retrieval-augmented generation (RAG). 

AI agents represent a new era of enterprise automation, predicted to be integrated into 33% of enterprise software by 2028. This shift could drive up to $4.4 trillion in global economic impact annually, with the market growing from $5.1 billion in 2024 to $47.1 billion by 2030.  

• Data quality matters: High-quality data ensures accurate agent performance.  

• Semantic layers are essential: Tools like LookML and BigQuery enable agents to understand relationships and context.  

• Adaptability is key: Observability and reinforcement learning improve agent performance in dynamic environments.  

• Governance is critical: Robust policies ensure data privacy and security.  

• Unified platforms simplify integration: Combining structured and unstructured data with AI capabilities accelerates innovation.  

My thoughts:  I expect we'll see more articles like this, especially after gardeners reporting a few weeks ago on expected increases and spend in the next year.   There are some good principles to follow but what is missing is a reference architecture for people to start adopting in some of these different areas.   Developing agents is  more complicated than a lot of the other generative AI projects people have been able to do in the last year.  Understanding how to harness the reasoning and planning component of a model is going to take time for teams and early on is only going to accomplish 2 to 5 steps worth of tasks until there is a deeper understanding of how these projects will need to solve business problems.  I still think people are going to find tremendous value early on, but I think the bar for developing these projects in the house is going to be much higher until people make that fundamental shift of going from single inference to multi step inference.   I actually look forward to talking to a few of my friends that work as CIO's and CTO's and some Enterprises to see how they are thinking about this with their teams.

AI agents could revolutionize travel planning and outdoor adventures. - (Lumb @ CNet)

Future AI agents promise to streamline trip logistics, adapt to changing conditions, and enhance personal safety, making outdoor experiences more enjoyable and efficient. During a hike to Haleakalā’s 10,023-foot peak, the author reflected on Qualcomm’s vision of "agentic AI." Announced at the Snapdragon Summit, these on-device AI agents aim to optimize daily tasks by integrating data from apps, devices, and real-time conditions.  

AI-powered tools like Qualcomm’s envisioned agents could redefine personal technology by offering hyper-personalized, real-time assistance across activities, bridging gaps in connectivity, and improving user autonomy in remote or complex environments.  

• Travel logistics: AI agents could create optimized itineraries, combining app data for food, weather, and routes.  

• Real-time adaptability: They could warn of closures, suggest detours, and check weather to adjust plans dynamically.  

• Offline functionality: Future agents may utilize satellite data to navigate and communicate when disconnected.  

• Safety tools: Agents could summon medical help or provide first-aid advice in emergencies.  

• Human collaboration: Despite advancements, human expertise, like park guides, remains irreplaceable for nuanced advice.  

My thoughts:  I thought this was just a fun piece where the author walked through how their trip to Hawaiian volcano would've been different. Had they had an agent to help them do so. If you are working on agents, then these are things that you're doing every day. But I think it's pretty eye-opening if you're just starting to learn about how agents work or could work and how they can affect you as a consumer but also the products you're building. Nothing in this article was science fiction, these are all breakthroughs which we should all expect to see in the near future. 

Honorable Mention, Open source breakthroughs

Cerebras Inference sets a new record: Llama 3.1-405B runs at 969 tokens/second, redefining frontier AI speed - (Wang @ Cerebras)

With unprecedented performance and minimal latency, Cerebras Inference makes frontier AI models viable for real-time applications, significantly enhancing user experiences across industries like voice, video, and reasoning. Cerebras has shattered performance barriers for large AI models, surpassing GPUs and other solutions. By optimizing Meta’s Llama 3.1-405B, it delivers the fastest output and lowest latency, revolutionizing AI deployments.  

This breakthrough positions open-source AI models as cost-effective, high-performance alternatives to closed systems, accelerating adoption in applications requiring speed, long context, and reasoning capabilities.  

• Speed: 969 tokens/second—12x faster than GPT-4o and 18x faster than Claude 3.5 Sonnet.  

• Latency: Time-to-first-token at 240ms, drastically improving real-time interactions.  

• Long context: Handles 128K context length at record-breaking speeds.  

• Pricing: $6M/input token and $12M/output token, 20% cheaper than AWS, Azure, and GCP.  

• Availability: Customer trials available now; general release in Q1 2025.  

My thoughts: It is not every day that you can accomplish breakthroughs on both speed and cost. This is such a significant breakthrough on speed.  I really just wanted to take a moment to share this because I saw it as an exciting breakthrough and can't wait to see what comes out of this group next.