Having seen and done my bit in the agentic AI field in 2024 with knowdroids.ai, I dare to share a couple of educated guesses about what might happen in the next 12 months. Well, a guess smells a bit like a game of chance; I am sure some of the items below are different from a guess then - first, they are opinionated, and second, they often are on someone’s roadmap, and that someone will strive to make them happen. Still, I may be proven wrong at the 12-month horizon. For the comfort of structure, I’ve put the topics into the categories of Models, Agents, Adoption, and Regulation.
Reasoning-capable models, such as o1 or Qwen, will come further to prominence and will get trained or fine-tuned not only on datasets like Math, Physics, Biology, or Chemistry but also on business-reasoning datasets. These are yet scarce at the moment, but in some business fields - e.g., in law via judicates - they are within reach.
Further, closed LLMs from O.AI, Anthropic, Google, etc., will still play their prime role and be state-of-the-art. However, the Open Source field will be in close tandem, with OSS-driven applications flourishing at very affordable total costs and offering better knowledge privacy than the closed world can offer. Multi-model applications that draw the best from both worlds will become a distinguished category.
They are already one of the main promises of AI today and in 2025 they will prove themselves to be the killer app in business settings. Many prior AI applications - such as Machine Learning models - will be wrapped into agentic tools, using the anthropomorphic character of agents and enjoying the resulting lower barrier for adoption.
Single-task agents will routinely get ensembled into complex process agents, collaborating across business functions and sectors. AI Agents of the world will gradually unite.
Agents will work the smoothest if all kinds of APIs undergo further standardization to become widely utilizable tools. Additionally, inter-agent communication (even for heterogeneous agents) needs further elaboration to allow for interoperability - another field to be paid attention to in 2025.
Many of the agentic solutions of present time are in a sense “static” - not improving much over time, besides perhaps their LLM brains being potentially replaced for newer models or storing their history of interactions as a long-term memory - new, self-improving or self-evolving paradigms will enter the scene.
Thanks partially to the anthropomorphic nature of autonomous agents (Look! The thing talks to me!) and partially to their ability to increase business productivity, AI penetration and adoption will accelerate in 2025. A little later - my estimate is though not sooner than in 2026 - when agentic technologies mature into really simple-to-design/produce/deploy tools at an educated user’s wish, I expect a moment of agentic Cambrian explosion and their ubiquity.
I’ve heard estimates of the cost of EU AI Act implementation - counting in percentage point(s) of GDP. Then I see the steeply widening gap between European and US (and Asian) productivity and the dismal pace of European innovation. And then there’s the unstoppable trend of aging in practically all European economies.
I can’t stop thinking that something will have to give here - and that in the context of an aging population and fertility rates close to 1.00 in several countries, maybe “we need every helping hand, even if artificial” will prove to be a more useful starting point than the current EU regulatory paradigm. I just wonder whether a shift will happen already in 2025, or a tad later.