Beyond the AI Hype: Why Building Reliable AI Agents Is More Difficult Than Building Intelligent Ones
Artificial intelligence is no longer just an emerging technology discussed primarily by researchers and technology enthusiasts. In the modern business landscape, companies are investigating the potential of AI to revolutionise customer service, automate processes, aid decision-making and enhance efficiency.
The most exciting advancement in this area is the emergence of AI agents that can not only respond to instructions but also perform tasks, interact with tools, retrieve information and complete complex workflows, with varying degrees of autonomy.
But as someone working on the engineering and infrastructure of several AI projects, I feel there is a crucial conversation that isn’t happening often enough. Creating an AI agent that appears intelligent is one thing. Creating one that businesses can consistently rely on is a whole other story. The extent to which organisations will successfully implement AI in the future will depend on the gap between the two.
We’ve seen some pretty impressive AI demonstrations. A developer presents an agent that can answer questions, write reports, retrieve information or automate a business process, and there’s an immediate reaction of excitement about what the technology can do. But a successful demonstration does not necessarily translate into a reliable production system.
How does an agent deal with incomplete information? What happens when an external service becomes unavailable? Is it able to recognise when it does not have sufficient information to make a decision? Can it recover when an operation is unsuccessful? More importantly, can its actions be examined and verified? These are not minor technical considerations. They are essential requirements for implementing AI in real-world business environments. Imagine a financial services firm in Nigeria using an AI agent to respond to customer enquiries. The agent might perform exceptionally well during testing, but if it provides incorrect information about a financial transaction or fails to escalate a sensitive request, the consequences go beyond a poor customer experience.
The same applies to a UK healthcare organisation, professional services company or financial institution. An agent that produces convincing but inaccurate information can introduce operational, financial and reputational risks. That is why I believe the industry needs to shift away from measuring AI progress primarily through intelligence and focus more deliberately on reliability, accountability and predictable performance.
Another issue in the AI landscape that isn’t as apparent is vendor lock-in. Many AI agents are built around the tools, APIs, memory systems and execution environments of specific AI service providers. This can result in an agent developed for one model provider requiring significant redevelopment before it can operate effectively with another.
This poses a challenge for businesses: what happens when the technology they rely on evolves?
A business can build its operations around a specific AI provider, only to find that another AI model is more effective, less expensive or offers capabilities that better meet the company’s needs. Ideally, transferring between providers shouldn’t require rebuilding the agent from scratch or losing the operational knowledge and performance history accumulated over time. Unfortunately, interoperability remains a significant problem. One of the issues I’m trying to solve with Wardscribe, my AI agent registry, is portability and interoperability.
With Wardscribe, I developed a structured approach to packaging, publishing, discovering and deploying AI agents across compatible environments. The system implements the Open Agent Schema Framework (OASF), an open interoperability schema published by AGNTCY, a project of the Linux Foundation. The goal is simple: an AI agent should not be permanently tied to the infrastructure on which it was originally developed.
This is significant because the future of AI should not be about standalone platforms that cannot communicate with one another. It should be characterised by systems capable of exchanging capabilities, integrating effectively and operating across different environments.
Portability provides greater flexibility for businesses when selecting technology providers. It also minimises the risk of building critical operations around infrastructure that is difficult or costly to change.
A key takeaway from developing AI infrastructure is that reliability cannot be assumed based on how well a model performs in a conversation. An agent can produce good responses but still fail when asked to perform a series of actions. For instance, an agent handling a business request might have to retrieve information, validate it, interact with an external application, perform an operation and communicate the outcome.
Each stage introduces potential failure points.
A reliable system must therefore take several engineering requirements into account:
- Clear boundaries and permissions
Agents should operate within the scope of their responsibilities. They should not automatically have access to every system, database or operation simply because they can interact with them.
- Verification and validation
Appropriate requirements should be used to check outputs and actions. If an agent cannot establish confidence in an outcome, it should be able to seek clarification or escalate the task.
- Version control and auditability
Businesses must know which version of an agent they are using, what has changed and whether it is possible to restore a previously tested version.
- Failure handling
An agent should be capable of identifying unsuccessful operations, handling interruptions and avoiding the presentation of incomplete tasks as completed.
- Performance monitoring
Organisations need a way to measure their agents’ performance over time, rather than relying on occasional demonstrations or subjective impressions.
These principles are especially significant as AI transitions from experimental initiatives to systems that support everyday business operations.
My work on Wardscribe has further convinced me that future generations of AI infrastructure need to solve problems that extend beyond model intelligence. I designed and developed the platform around several components intended to make AI agents more portable, inspectable and measurable. One of them is WardPack, a signed and immutable packaging format for AI agents.
Instead of treating an agent as a standalone set of code and instructions, WardPack provides a structured package that includes the agent’s configuration, identity, skill dependencies, tool input and output contracts, memory schema, workflows, provider-specific profiles and evaluation benchmarks. It also contains cryptographic signatures to support the verification of published packages.
The significance of this in practice is that an organisation can work with and audit a more complete and inspectable representation of an agent, rather than relying solely on the environment in which it was built. WardMind is another component: a confidence-scoring model that evaluates agent quality across eight dimensions. The reasoning behind this is crucial: an agent’s performance should not be judged solely by how well it generates a convincing response in a single interaction. It needs to be assessed more systematically. The confidence model is intended to produce additive and portable performance information, allowing an agent’s evaluation history to remain useful when it is moved to another compatible runtime. This is particularly relevant in an industry characterised by rapid changes in models, providers and deployment environments.
Wardscribe also incorporates search and discovery infrastructure, version management, signature verification and support for multiple API protocols, including REST, GraphQL and gRPC. These are not features designed simply to give the impression that the product is technically advanced. They address real-world issues surrounding agent distribution, verification, discovery and integration into existing technology environments.
This is where I think meaningful AI infrastructure development happens: identifying limitations that affect the broader ecosystem and engineering solutions that can be used beyond a single application.
The discussion about AI adoption cannot be confined to Silicon Valley technology firms or multinational corporations. In Nigeria and the UK, businesses are already exploring the potential of AI to boost productivity, minimise repetitive tasks and enhance customer experiences. But the priorities and challenges in these markets are not necessarily the same.
In Nigeria, many businesses operate in environments where technology budgets, infrastructure reliability, workforce capacity and access to specialised technical expertise often influence decisions about technology adoption.
For small and medium-sized businesses, it’s not just about whether AI can perform a task. It’s whether the solution is affordable, reliable and capable of integrating with existing business operations. For instance, an AI-powered customer support system can alleviate the burden on a customer service team. However, if it regularly misinterprets local customer enquiries, cannot integrate with existing business systems or is expensive to maintain, its overall value becomes questionable.
It’s important for businesses to focus on clearly defined use cases and measurable outcomes, rather than investing in AI simply because others are doing so.
In the UK, organisations might have more established digital infrastructure, but they still face challenges with system integration, data protection, governance and accountability. In regulated or sensitive industries, it is crucial to have visibility into an AI agent’s operations, including what data it accesses and how its decisions can be reviewed.
I think the same principle should apply across both markets: AI adoption should create operational value and be supported by solid engineering.
Businesses should consider whether an agent addresses a genuine problem, whether its effectiveness can be measured, whether its actions can be controlled and whether the organisation can switch providers without becoming unnecessarily dependent on one technology ecosystem.
With the continued advancement of AI capabilities, access to intelligence alone might no longer be a competitive advantage. The ability to build reliable systems around that intelligence will matter far more. More organisations are likely to experiment with autonomous agents, but not all these experiments will become successful business deployments. The organisations that succeed will be those that understand the importance of infrastructure, interoperability, monitoring, security and human oversight.
Developers also have a responsibility to approach this transition carefully. Creating an agent that can perform a task is just the first step. It is also important to consider what happens when things go wrong, how its performance is assessed and how its capabilities can be verified. That’s why open standards and collaborative infrastructure development are important. Common approaches that enable different systems to communicate, share capabilities and operate without unnecessary technical barriers will benefit the agentic AI ecosystem.
My work with Wardscribe reflects that belief. I see the future of AI agents as an ecosystem in which portability, transparency and measurable performance are built-in design considerations rather than optional features.
The hype around artificial intelligence is understandable. The technological landscape is undergoing a significant transformation, and AI agents are poised to reshape how organisations approach work, services and automation.
However, the true measure of progress is not how impressive an agent looks during a demonstration. It is whether it can work consistently, stay within its defined boundaries, recover when things go wrong, protect information and deliver results that people and organisations can count on.
As AI continues to become part of everyday business operations, this distinction will grow in significance for businesses in Nigeria, the UK and beyond. I believe the next stage in the evolution of AI belongs not only to those building more intelligent models, but also to those working on the difficult problems of making intelligent systems reliable. Intelligence gets attention, but reliability earns adoption. Without trust, autonomy has very little business value.
Written by Daniel Babalola


