Dave: Back in August last year, you wrote an analysis of Agentic AI focused on orchestrated automation workflows that combine LLMs with specialised tools. We've since seen rapid developments in multi-agent architectures and human-AI collaboration frameworks. Could you explore how your initial predictions about agentic flows have evolved, particularly regarding reliability and governance?
Justin: I was pleasantly surprised to see that much of what I wrote holds true, despite the rapid advances in models and more real-world implementations of agentic AI. It’s no surprise that large platform providers like Salesforce and Google now provide agentic workflows and configurable agents for common on-platform tasks through offerings like AgentForce and AgentSpace.
DeepSeek’s advancement in delivering a reasoning model with lower training and inference costs should boost more sophisticated use cases. The advancement in the integration of real-time data and content sources is also one to watch, and 2025 looks promising for the development of AI-first products and services, rather than AI-augmented ones, with agentic-based systems likely to take the lead.
There are still questions over ethics and governance that have to be addressed as these systems come online and become more complex and autonomous, particularly when we start entering the territory of giving true agency to them.
Reliability is definitely a challenge when developing and running these types of systems: we are using observability and monitoring tools, and giving access to these to a set of validate and control agents as part of the systems we are developing. This is proving to be a successful approach, which also addresses the issues of transparency and accountability - something we should all be looking for when deploying this kind of solution in real-world situations.
Dave: What patterns are you seeing in how enterprises, especially in our core sectors of research & insight, training & learning, healthcare and publishing, are approaching the balance between automation and human oversight?
Justin: At Spicerack, we’ve been developing agentic reporting and strategic analysis solutions that are multi-agent with human-in-the-loop inputs during the reasoning, research and review phases of the typical content ideation, creation and publishing process. This is a good example of AI/Human collaborative work, which enables people to focus on the high value and creative aspects of the content creation process, which ultimately delivers unique and non-formulaic content.
A common-sense approach is required based on the scenarios in which we deploy Agentic solutions. In healthcare and finance applications, I expect more human control loops than I might for a simple content summariser. Ultimately, humans will be the accountable ones for a while longer, so we should optimise the human/AI interface and be clear on where we want the choice and control boundaries to be.
