By now, you’ve probably seen a ton of articles about AI agents in the news, whether it’s new systems being released or millions of dollars in funding being raised. Money is flowing into agent startups like crazy, and even many of the largest companies like Microsoft and Salesforce have been putting a lot of effort into open sourcing, operationalizing, and commercializing agent-based systems that leverage some of their investments in LLMs. It’s clear that agents and multimodal systems are going to be a massive part of the next phase of AI and that they won’t be moving out of the zeitgeist any time soon.
Recently, Arthur’s co-founder and Chief Scientist John Dickerson hosted a session where he dove into the world of agents—exploring their functionality, use cases, and how they came to be. If you’re interested in learning more about the history of agents and how they work from a technical perspective, certainly give John’s session a watch. Or, keep reading to find out what he predicts is in store for agents in 2025.
1. Cost will matter more than ever.
These are not free systems to use—and agentic AI is even more expensive than traditional LLM-based applications. At Arthur, we’re already hearing tons of questions from CFOs and CIOs about measuring ROI of basic LLM-enabled applications. If 2023 was the year of the science project and 2024 was the year of the production product, John predicts 2025 is going to be the year of looking at your wallet. For public companies, shareholders will be looking at capital expenditure, and for private companies, CEOs and CFOs will be paying close attention to spend as well.
2. Systems of multi-agent systems will emerge.
Individuals are already using LLMs, single-agent systems, and increasingly multi-agent systems to perform individual tasks. However, these fleets of agents don’t interact in a vacuum apart from other people’s fleets of agents—so it will be crucial to understand how to orchestrate all of this. True AI-human teaming occurs when you have systems of humans enabled by their own individual multi-agent systems working together.
3. LLMs and agents will be used as social surrogates.
In areas like market research and user testing, multi-agent systems are being used to replace humans. Even Microsoft recently launched TinyTroupe, an “LLM-powered multi-agent persona simulation for imagination enhancement and business insights,” with use cases such as advertisement (simulated audience evaluation) and brainstorming (focus group simulation). If you’re interested in reading a counter-argument that outlines the risks of this approach, our team actually published a paper about just that.
4. As always, there will be snake oil.
There’s a lot to be excited about when it comes to agents, and the combination of planning/reasoning approaches with present-day LLMs is quite powerful. That said, there is so much venture capital, private equity, and FAANG money being poured into agentic AI right now—it’s inevitable that some nonsense will come out of it. The key to identifying the diamonds in the rough? Understand your firm’s needs, do a cost-benefit analysis of agentic systems, and understand risk measurement and mitigation.
Interested in learning more about agents? Register for John’s next webinar here, and check out Arthur’s agentic support capabilities as well.
SHARE
