AI Agents are the new shiny toy – Here’s where to actually start
AI agents are the new shiny toy, and it feels a lot like the LLM hype from two years ago. Plenty of ‘experts’ are now touting agents as the endgame of AI, the same way they once touted LLMs as the dawn of sentient machines.
I recently read an article claiming a wave of unicorn AI startups with no employees at all. If you believe that, I’ve got a bridge to sell you in New York.
To be clear, I’m genuinely bullish on AI agents, but only where the agentic workflow and the underlying cognitive architecture are implemented correctly. The harder problem, and the one that actually determines success, is integrating agents cleanly into your existing enterprise architecture. That challenge doesn’t go away even for teams willing to start from scratch.
If your architecture already runs on microservices, the fastest entry point is wiring your existing APIs and services in as ‘tools’ for the agents via function calling. That alone gets an agentic workflow started without a rebuild.
Here’s the CxO-relevant starting framework – begin with anything that already runs unattended. Nightly batch jobs are the cleanest first candidate, along with data- and information-gathering tasks that feed a decision, like support-ticket triage. These are low-risk, high-repetition, and forgiving of early mistakes. Exactly the app profile you want before agents touch anything customer-facing.
The trick, as it always is with a new capability, is to start small and expand the footprint deliberately rather than announcing an ‘agent-first’ strategy before you’ve proven the first one works in production. If you’re past the batch-job stage and mapping out where agents go next, that’s the conversation worth having before the roadmap gets locked in.