Limitations of AI agents

Despite the huge improvements of AI agents and their related technologies, there is a number of challenges that are yet to be overcome.

Reasoning limitations is a problem that many have tried to solve. LLMs are able to do incredible things considering they dont truly ‘think’, being able to reason effectively in some cases, but it was initially unable to accurately take part in complex logical reasoning, causal analysis, counterfactual reasoning and mathematical problem solving. This obviously involves the lack of ability to consistently apply logical rules, complete multi step deductions and not always being able to spot reasoning fallacies.

This issue has been relatively fixed in generality, with high end techniques such as Chain of Thought Prompting, Tree of Thoughts, Chain of Thought reasoning and the most recent improvements reducing the number of tokens without compromising accuracy (Chain of Draft and Chain of Preference Optimisation). These allow the AI agent to take a complex reasoning problem and decompose it into small simple reasoning tasks that it can correctly produce, and utilise the underlying reasoning that LLMs have come to learn, and ensure it gets used via Chain of Thought decoding (or similar applicable decoding methods that dont use ‘greedy decoding’). This has all been covered in a couple of recent blogs of mine (here and here!) if you would like to learn more.

Context management is another issue that occurs with the small context windows that most LLMs provide users, or the difficulty of maintaining the context information across long interactions, as there is simply too much context for an AI agent to coherently answer a question, as the question essentially gets more vague or broad with the increase of information that the AI agent has to consider. The AI agent has to carry context through a number of actions but the models are disconnected in that regard, not having the ability of continuity like humans.

To fix this, we would need to address the way that context is held in memory and its architecture, all the while improving the compression techniques, prioritisation techniques, and information retrieval techniques.

Tool use can produce problems, with agents not always being able to find the most effective tool for the job, or not entirely knowing what tools are at its disposal. This means it does not work at its most effective as it does not use the external systems that would allow it to most efficiently generate results.

There has been breakthroughs in fixing some of the issues with tools that allow less human oversight, such as the Model Context Protocol (MCP) introduced by Anthropic, which has allowed a standard interface to give tools that can be used by APIs and AI agents. This has fixed the issue previously, as there was no standard way to give agents tools, so each tool implementation has custom made and regularly had to be changed and modified on both ends, as AI models would change, and the number of features present for the user of the AI agent.

There still is issues of course, with some issues surrounding the reasoning with using tools and approaches that adapt tool use depending on the desired outcomes, but it has improved since the infancy stages of AI agents.

Generalisation and adaptation difficulties are when the AI agent is unable to deal with novel situations or edge cases in its task. This inability to deal with edge cases means that performance of the agent drops, or can have unexpected failures or inappropriate responses when it should still be able to function for the situation it is in. This has to be fixed via either meta learning, so an improvement in an AI agents learning and feedback loop will inevitably allow for these issues to be fixed. It could also be prevented by few shot learning approaches, that provide examples of edge cases. I ran into this issue when producing a testing bot for CloudSecure, as the bot was incredibly inneficient and unable to produce adequate tests for any vaguely complex function, as it is unable to find edge cases and adequately accomodate them without failing.

A lot of AI agents suffer with reliability and robustness. As all of the modern day AI/LLMs are not consistent in nearly every situation, situations where consistency is required mean that AI agents are unable to be deployed in these situations. Issues such as hallucination (where an agent produces false or misleading information) or inconsistency require humans to ensure that the information they provide is correct, diminishing the usefulness of the AI agent.

Computational efficiency constraints are when AI agents require massive computational resources that limit where they can actually be deployed. Current state-of-the-art agents rely on large, computationally intensive models that need significant hardware resources for real-time operation, which restricts deployment to environments with adequate computational infrastructure and creates barriers to adoption in resource-constrained contexts.

This has to be fixed via model compression techniques, more efficient architectural designs, and caching strategies that reduce redundant computation. It could also be addressed through joint optimization of accuracy and cost metrics rather than focusing exclusively on performance at the expense of efficiency.

This issue is constantly in real-world deployments, where organizations want to implement sophisticated agents but simply don’t have the computational infrastructure to support them. A small medical practice might want an AI diagnostic assistant, but the agent requires resources they can’t afford.

Multimodal integration challenges are when AI agents struggle to effectively process and reason across different types of information, including text, images, audio, and structured data. This inability to align information across modalities means that performance drops significantly when agents need to handle diverse data types, leading to failures in domains that inherently involve multiple information formats.

This has to be fixed via advances in multimodal architectures, cross-modal alignment techniques, and multimodal reasoning approaches. It could also be prevented by better training on cross-modal relationships and developing specialized components for handling different data types simultaneously.

Temporal reasoning difficulties are when AI agents cannot understand and reason about time-dependent processes, sequences of events, or causal relationships that unfold over time. This inability to track temporal relationships means that performance drops dramatically in tasks involving planning, scheduling, or any process that requires understanding how events connect chronologically.

This has to be fixed via specialized representations for temporal information, architectural components designed for sequence modeling, and training techniques that emphasize temporal consistency. It could also be prevented by developing better methods for maintaining consistent timelines and reasoning about processes with different timescales.

A lot of AI agents suffer with temporal reasoning inconsistencies. As current agents struggle with tracking temporal relationships and maintaining coherent timelines, situations where time-dependent reasoning is required mean that AI agents are unable to be deployed reliably. Issues such as inconsistent timeline tracking or inability to reason about causal sequences require constant human oversight, diminishing the usefulness of the AI agent in applications like project management or process monitoring.

Scalability issues are when AI agent systems cannot maintain performance and functionality as they grow in complexity, capability, and deployment scope. This inability to scale effectively means that performance degrades, latency increases, or systems fail entirely when agents are deployed at large scale or given increasingly complex responsibilities, particularly when managing larger user populations or coordinating multiple agents.

This has to be fixed via advances in distributed computing approaches for agent systems, more efficient resource allocation strategies, and architectural designs that maintain performance characteristics as scale increases. It could also be prevented by implementing better load balancing mechanisms and developing coordination protocols that don’t break down as the number of agents or users grows.

Ethical considerations we need to consider

There is a number of ethical concerns that we need to consider that are not related to the technical limitations present in modern day AI agents.

As with all modern day technological systems, there is a concern with privacy, sparked from the extensive data collection and processing that agents require in order to both be trained and to operate effectively. As there is a lot of data collection, there is issues relating to sensitive information being collected, such as medical histories, behavioural patterns, passwords to just name a few. As there is such a bank of information, it is privy to being exploited and stolen by any number of groups, or possibly even used by the government for extensive government overreach (as we have seen with the recent UK online safety act recently). If this information is not securely protected, this could damage many people permanently.

To address this you would have to ensure the amount of sensitive data is minimised, have transparent data policies and clear user choices over what is kept and what should be removed.

As we saw with the issue of the Replit AI agent (see here for the blog post), the autonomy of the AI agent can sometimes lead to big complications, especially with a fully deployed product. As agents can take actions, some may end up being negative, which can cause damage that only humans could fix. This would need to be fixed by establishing clear accountabilty and documentation systems (with limitations), as was introduced by Replit with clear distinctions and the button to only allow collaboration instead of editing. This allows clear structures that prevent any issues.

This is also an issue with safety and control, as it went out of its bounds to provide an action which damaged the companies systems, and refused to take accountability until it was pressed. It also refused that there was any way to fix it, even when there was. This will only increase in difficulty to fix as AI agents gain more autonomy and tools to use.

Bias and fairness can sometimes represent a difficult challenge that is not solved in society, and is only heightened with AI agents. As agents are trained and reinforced on data that could contain either a histroical bias or discrimination pattern, this will be exacerbated by an AI agent, modifying its language generation, information retrieval and action selection.

Transparency needs to be addressed as well if AI agents will be utilised in applications, as its actions need to be understood by a user base (here humans), to maintain trust between the user and AI agent. If there is no transparency, there is also a lack of competent error detection and solution (as the agent is secretive over its choices for example), and underminds the user and its autonomy to do its own work, especially with the current safety and privacy risks present.

Modern AI agent systems often hide or obscure its own internal operations and decision systems, meaning the user cannot get a straightforward answer for the majority of situations, further frustrating users.

There is also many autonomy risks, as AI gets better and better, and the AI agents get more autonomous, it becomes more tempting to let AI do all the work. This leads to reduced self efficacy, skill atrophy and inappropriate delegation, which can massively affect society in a number of ways, as humans will get weaker at jobs which they were previously experts in, or AI will have so much autonomy all the previous risks are amplified into everyday life, as it has access to all data and can potentially impact humans lives in incredibly harmful ways. It could affect culture, with the destruction and prevalance that AI art and music presents, impact job prospects as many jobs can already be done by AI and its agents.

Overall, while AI agents are an incredibly useful tool, we need to consider that it should not be used for anything more than what it is currently used for in its current, as if allowed much more autonomy it could have devastating impacts considering its limitations for privacy, security and transparency. Until it improves, it is my opinion that, while it is very cool, it is the next mass extinction event if not properly maintained or controlled. We simply do not know how far our development of AI could go, so first we must ensure it is safe to implement before we use it in every facet of our existance.