It’s the season of the pundits – everyone with a megaphone or a Medium/Substack/LinkedIn handle has to have an opinion on what 2025 will be for the rest of us lesser tech mortals. We have seen plenty of posts about this in the last few weeks, and almost all of them seemed to say only the same thing – 2025 is the year of the AI agents! I didn’t want to sound contrarian or use a clickbait title to tell you the exact opposite. I have been patiently soaking in all the noise around me for the last year or more and even dipping my toes to test the waters. Based on my observations, experiences, and peer conversations, 2025 may NOT be the Year of the AI agents. Humor me and read more to understand my reasoning better. I am not saying AI agents are the wrong technology choice for this year. Like it or not, we are already in the age of agents. My point is that agents don’t have the legs to stand on tomorrow.
The Definition Problem: Not Everything Is an Agent
The first challenge lies in the very definition of AI agents. The industry has developed a disturbing habit of labeling any AI-powered interface as an “agent,” regardless of its actual capabilities. Traditional chatbots and simple automation tools are being rebranded as agents without meeting the fundamental criteria that define true AI agency. Three months ago, I talked with a startup founder who claimed he has an “AI agent platform” for recruiting (get in line behind the 1000 other ones out there!) Once he showed me the demo, I realized that he was just showing me a SaaS product. I asked him where the AI/agent was. He said that it was behind the scenes, and his explanation made it look like a bunch of business rules. It doesn’t take much to realize that people are truly AI-washing/Agent-washing everything to score some attention points, but you can sniff them out from a mile away.
An AI Agent must meet certain key requirements (but not limited to) such as –
- Autonomous goal-directed behavior
- Persistent objectives and state management
- Strategic planning and adaptation capabilities
- Meaningful environmental interaction
- Independent decision-making abilities
Even at a recent mega-event of one of the biggest software companies in the world, they announced over 40+ “agents” when, in reality, they were more like co-pilots than actual agents. No autonomy in any of those “agents”! The loose application of terminology going around the industry creates unrealistic expectations and undermines serious discourse about AI agent capabilities and limitations. This can be a detriment to the success/failure of an “AI agent” deployment.
The Accuracy Challenge: When 99% Isn’t Good Enough
While current AI systems demonstrate impressive capabilities in controlled environments, the accuracy requirements for mission-critical enterprise applications remain beyond their reach. In domains such as healthcare, banking, and financial services, even a 1% error rate is unacceptable.
The accuracy challenge becomes even more pronounced in multi-agent architectures or multi-step agentic workflows. When multiple agents work in sequence (or orchestrated), errors compound and cascade through the system. Consider a workflow where three agents process data sequentially, each with a 95% accuracy rate. The cumulative effect of these cascading inaccuracies can result in significantly compromised output quality. This “error propagation” effect makes multi-agent systems particularly risky for mission-critical applications. Until we solve the accuracy problem, it is premature to get into a “Year of AI agents” celebration. Also, note that I didn’t even mention the pink elephant in the room – Hallucinations! That is a series of blog posts by itself.
Security and Data Privacy: A Critical Gap
Those following me know I am passionate about this space and have even written about it recently. Enterprise deployment of AI agents faces a fundamental security challenge. Current AI systems struggle with consistent identification, categorization, and protection of sensitive information such as PII (Personally Identifiable Information) and PHI (Protected Health Information). This limitation creates significant risks –
- Potential breaches of regulatory compliance (HIPAA, GDPR, etc.)
- Inconsistent handling of sensitive data across interactions
- Risk of indirect information leakage through inference
- Challenges with data residency requirements
- Inability to maintain consistent security boundaries
Until these security concerns are adequately addressed, enterprise adoption of AI agents will remain limited to non-sensitive applications. The more significant challenge here is that we truly don’t know where PII/PHI data are stored anymore, given the advent of so much unstructured data across enterprises. Earlier, if we designated a table column to hold sensitive data such as SSN or customer address, we would know how to mask or protect it from unauthorized access. But, now, such sensitive information is rampant across unstructured text and files everywhere, and even the security teams are mostly unaware of them. And when LLMs get hold of such unstructured text/files, they are not any smarter either to detect or flag them. At least, not very accurately and consistently, even if they did.
Agents are NOT taking over the world – at least, yet!
In the last few months, we’ve seen this trend where, as soon as a new feature is announced by one of the larger AI companies (Meta, OpenAI, Google, etc.), the “pundits” start immediately proclaiming that an entire industry is dead. For example, look back at the number of “RPA is dead” articles that came out when the LLM companies started announcing beta features that did some elements of enterprise automation like data extraction from images, mimicking keyboard/mouse clicks on your computer screen etc. While these are commendable evolutions with LLMs, they are hardly accurate or consistent to put them into a production system and replace your RPA solution. At best, LLMs and agents are augmenting RPA or other automation solutions out there today. Without a human-in-the-loop, it is hard to convince yourself that you can trust the agent to do everything by itself today. We will get there. Some day. Just not yet.
Which company has completely replaced its developers with agents? Which programmer or DevOps lead, worth their salt, would trust an agent to autonomously code, test, and deploy the code into production today? None of that has happened, yet hundreds of articles and posts claim that companies will hire agents instead of programmers. Again, this may happen in the future, but we are still far from it.
What to expect in 2025
I am still not going to predict anything as we live in these crazy and exciting times, but I am going to tell you what you should expect – more like demand from these AI companies.
- Better accuracy and reliability – We need agents to be more robust and consistent with anomaly detection, handling edge cases, and managing uncertainty in decision-making.
- Better security and compliance – Agents must have heightened sensitivity to data privacy requirements and be more aware of active and upcoming compliance regulations. They should be better at detecting, classifying, and protecting sensitive information from leaking into the wrong hands. AI guardrails is an important space and should evolve to support these needs.
- Better enterprise integration – While this may be handled with additional product augmentation, it is better that the agent itself is able to handle improved integration with existing workflows, handle legacy systems, and support robust audit and control mechanisms.
- Better multi-agent coordination – As accuracy improves, Multi-Agent Systems (MAS) need to enhance their inter-agent communication, manage error-propagation, and improve task coordination across the workflow.
Several amazing areas, such as agent memories (short-term and long-term), multi-modal agents, context-aware security, domain-specific agents, etc., are being worked on actively. We are going to see even more exciting innovations come out in 2025. I truly cannot wait to see what this year will reveal in this space. Meanwhile, stop hyping up things and stop getting hyped by everything that screams agents! We are in this together.