ONE-LINE STING: An agent takes actions. Nobody told you that. Nobody told you which actions. Here's what you actually need to know.
"They've given me an Agent. I don't know what it does. Nobody knows what it does."— Office Man
AI Agents: What Are They Actually Doing?
May 2026 · Issue One
WHAT IS IT?
An AI agent is an AI that does things, not just says things. Standard AI tools answer questions — type a prompt, get a response. An agent goes further. You give it a goal, and it figures out the steps: searching, writing, sending, updating, checking. It acts in a sequence until the task is done. You don't click each step. It runs the loop.
The word "agent" is now attached to everything, which is annoying. The useful question isn't "is this an agent?" but "what can it access?" An agent connected to your email, calendar, and files can do something real. An agent that only reads text is just a fancy chat window.
WHY NOW?
The underlying AI got good enough. For an agent to be useful, the model running it needs to reliably follow multi-step instructions, recover from errors, and know when to stop. Until recently, that reliability wasn't there. Now it mostly is — for well-defined tasks.
So vendors moved fast. Microsoft, Salesforce, ServiceNow, Google — all launched agent products in 2025. "Agentic" landed in every roadmap. Organisations started deploying before they'd decided what the agent was allowed to do. That gap — between capability and governance — is where most of the current confusion lives.
THE HUMAN ANGLE
Agents are changing what delegation means at work. You used to delegate to a person. You gave context, trusted judgement, had a conversation. An agent needs the task broken down differently — clear inputs, defined outputs, explicit permissions.
That's not easier. It's a different kind of work. Someone has to define what "done" looks like. Someone has to review the output. Someone has to take responsibility when it goes wrong. Agents don't remove accountability — they move it earlier in the process, to the person who set the goal and set the guardrails.
The right question to ask: "What happens when the agent gets it wrong?" Know this before you deploy anything consequential.
WHAT TO DO
Ask what your agent can access — that's the honest measure of its capability. Start it on tasks with clear right answers: booking confirmations, data lookups, draft documents for human review. Build in a checkpoint before it does anything irreversible. And read the documentation for the specific product you've been given, not general articles about agents — the details vary enormously.
The goal isn't to understand the technology. The goal is to know what it can do for you on a Tuesday.