AI Agents Are Coming. But Are They Actually Useful Yet?

The idea of autonomous AI agents - digital entities that can plan, reason, and act across tools without human oversight - sounds like science fiction finally arriving. The pitch is seductive: agents that book your meetings, answer customer queries, build prototypes, write code, and optimise systems without you lifting a finger.

But if you’ve tried using them in practice, especially in complex enterprise settings, you’ll know that reality isn’t quite there yet.

At New Icon, we’ve been watching the rise of agentic AI closely. We’ve even begun integrating controlled agent behaviours into some of our products. But let’s be honest: most current-gen agents are a bit like enthusiastic interns with no memory and too much confidence.

Here’s why.

1. Short-Term Memory, Short-Term Thinking

Despite rapid progress in LLMs, most AI agents struggle to retain context over time. Planning a multi-step task? Setting a long-term goal and adjusting as conditions change? That’s still a tall order.

Agents today tend to operate in short bursts of context. They can handle individual steps such as writing an email, fetching a file, summarising a doc, but struggle to link those steps together in a meaningful, adaptive way. In a live system, especially one used by real users or customers, that fragility becomes dangerous.

At New Icon, we’ve found success using modular, goal-oriented workflows that simulate agentic behaviour but with tightly scoped memory and state logic we can control and test. That’s a far cry from “general purpose” agents, but it works and it’s safe.

2. The Tooling Trap

For agents to act, they need access to your tools: APIs, databases, CRMs, internal dashboards, scheduling software, and more. But enterprise systems aren’t plug-and-play. They’re often undocumented, siloed, or full of edge cases even your developers forgot about.

Without robust integration infrastructure and critically, guardrails, even the smartest agent will fail to execute anything meaningful. That’s why we often advocate for embedded intelligence, not just bolted-on LLMs. It’s one thing to connect an agent to a tool; it’s another to make sure it knows what not to do with it.

3. No Feedback, No Learning

One of the biggest challenges we see is the lack of structured feedback. In traditional software, you have logs, tests, monitoring, version control. With agents? It’s more like: prompt goes in, magic happens (or doesn’t), and you’re left wondering what went wrong.

We’re strong believers in testable systems. So when we prototype anything agent-like - a triage bot, an automated assistant, even a rules-based decision engine - we build in observability from day one. Feedback loops aren’t optional; they’re what separates a clever toy from a reliable tool.

So, Where Does That Leave Us?

We’re probably in the “dial-up internet” era of AI agents. There’s massive potential, but also huge gaps in capability, infrastructure, and real-world safety.

We don’t think the solution is to wait until agents are perfect. We think the solution is to constrain them thoughtfully - to pair LLM intelligence with well-architected systems, embedded testing, and a deep understanding of the problem space.

In other words: don’t let your agent freelance. Give it a job description, a sandbox, and a manager.

The Takeaway

We’re excited by the future of agents but we’re also wary of the hype. The real winners here won’t be the teams that chase autonomy for autonomy’s sake. They’ll be the ones who build patiently, modularly, and with an understanding that even the smartest agent still needs a system that supports it.

The agentic future isn’t cancelled. It’s just not fully hired yet.