From rule-based algorithms to agentic AI

When you ask Google Maps for the fastest route to a restaurant, nobody at Google plans the journey for you. When a suspicious prize draw lands in your spam folder, nobody at Gmail read it first. For decades this was what automation meant: software following rules written in advance, doing one job well and nothing else. The spam filter could not plan a route. The route planner could not read an email. Neither could learn a new trick without an engineer shipping it.

Large language models broke that pattern. Trained on billions of words, they could answer questions, draft documents, translate, and help debug code, all through plain conversation. One system could suddenly do many jobs. But the first generation had sharp limits. Their knowledge stopped at a training cutoff, they could not act on the world, and they could not check anything new without a person wiring them to it.

Agentic AI is the next step: systems that plan, decide and act with far less step-by-step instruction. An agent that notices traffic building before your meeting does not just warn you. It suggests an earlier departure, books the ride, and messages the people waiting. In a hospital, an agent can watch patient data against the latest studies and flag a doctor when something looks wrong.

The early deployments cluster where work is high-volume and structured. Customer service agents now resolve the majority of routine queries on their own. Security teams use agents to scan network activity continuously and surface intrusions faster than human review could. Supply chains, drug discovery pipelines and manufacturing lines are picking up agents that order, schedule and adjust without waiting to be asked.

The caveats are real. Agents make mistakes, their reasoning can be hard to inspect, and in hospitals and courtrooms those properties are not acceptable quirks. Many organisations trialling agents report exactly this worry. The direction of travel is towards multi-agent systems, where separate agents coordinate on traffic, energy and public services across a city, and eventually towards systems general enough to raise harder questions about safety and control.

The story from Google Maps to agentic AI is a story about software gradually acquiring initiative. Rules did one job. Language models did many, when asked. Agents act before being asked. Making that initiative reliable and accountable is now the interesting part of the work.