AI agents have been the technology story of 2024. The promise: autonomous systems that can complete complex tasks with minimal human oversight. The reality is more nuanced.

I’ve been tracking AI agent deployments across industries. Here’s what’s actually working.

The Agent Landscape

AI agents differ from traditional AI in a crucial way: they can take actions, not just provide outputs. An AI agent doesn’t just tell you what to do—it does it.

This sounds like science fiction, but practical implementations exist today:

Customer service agents: Handling inquiries, processing requests, escalating when needed. The best implementations resolve 40-60% of queries without human involvement.

Development agents: Writing code, running tests, fixing bugs. Tools like GitHub Copilot are the visible tip; more sophisticated agents handle larger development tasks.

Research agents: Gathering information, synthesizing findings, producing reports. Particularly valuable for repetitive research tasks.

Process automation agents: Moving beyond simple RPA to handle exceptions, make decisions, and adapt to changing conditions.

Where Agents Excel

AI agents work best when:

Tasks are well-defined: Clear inputs, clear success criteria, predictable scope.

Mistakes are recoverable: Low-stakes decisions where errors can be caught and corrected.

Volume justifies investment: Repetitive tasks that occur frequently enough to warrant agent development.

Human oversight is practical: Someone can review agent actions, at least initially.

Where Agents Struggle

Current limitations are real:

Novel situations: Agents handle variations of trained scenarios well but struggle with genuinely new problems.

Multi-step planning: Complex tasks requiring many sequential decisions with dependencies remain challenging.

Context switching: Moving between different types of work within a single workflow.

Explaining decisions: Agents often can’t articulate why they made specific choices, complicating troubleshooting.

Who’s Building Them

The AI agent ecosystem includes:

Platform providers: OpenAI, Anthropic, Google building foundational agent capabilities.

Enterprise software: Salesforce, Microsoft, ServiceNow embedding agents in existing platforms.

Specialists: A Sydney-based firm like Team400 creating custom agents for specific business applications.

Open source: AutoGPT, LangChain, and others providing frameworks for agent development.

For most businesses, the question isn’t building agents from scratch but choosing how to access agent capabilities—through platforms, specialists, or frameworks.

Practical Implementation

If you’re considering AI agents for your organization:

Start narrow: One specific, well-defined use case rather than broad automation.

Measure carefully: Track agent performance against human baseline. Be honest about results.

Plan for failure: Agents will make mistakes. Build in oversight and correction mechanisms.

Consider the humans: Agent introduction affects jobs. Think through workforce implications before deploying.

The Investment Question

Agent development costs vary enormously:

• Configuring existing platform agents: $10,000-$50,000
• Custom agent development: $50,000-$500,000+
• Enterprise-scale agent deployment: $500,000-$5,000,000+

ROI depends heavily on task volume and current labor costs. The business case is clearest for high-volume, well-defined processes.

What’s Coming

Agent capabilities are improving rapidly. Current limitations—planning, reasoning, reliability—are active research areas with meaningful progress.

Within 2-3 years, I expect:

• More reliable multi-step task completion
• Better integration with existing business systems
• Improved ability to handle exceptions
• Clearer frameworks for human-agent collaboration

The question isn’t whether agents will transform work—it’s how quickly and in what sequence.

My Take

AI agents represent genuine progress in automation capability. They’re not replacing human workers wholesale, but they are changing what work looks like.

The winners will be organizations that find the right human-agent balance: using agents for tasks they handle well while preserving human judgment where it matters.

That balance is still being discovered. We’re early in this transition.

---

Tracking the evolution of AI agents and autonomous systems.