AI agents: what to actually expect, and what is still coming
The word agent covers everything from a workflow with one AI step to a system that works unattended. It is worth separating what already works from what is still a promise.
Short answer
What works well today are narrow agents: specific tasks, with limited tools and a person reviewing the result. An agent that classifies incoming email or summarizes meetings and creates tasks is reliable. What still does not work is the autonomous agent that works for hours without supervision: it accumulates errors, drifts away from the goal and the cost takes off without anyone noticing.
- Last verified
- June 30, 2026
"Agent" is currently the word attached to almost anything with AI inside. It is worth separating things out, because the difference between the two ends of the range is enormous.
What already works
Classifying and routing. Incoming email, tickets, forms. The agent reads, decides the category and triggers the corresponding action. If it gets it wrong, you fix it by hand and nothing happens.
Extracting data from documents. Invoices, contracts, reports. Pulling specific fields out of a PDF is a task current models are very good at.
Summarizing and creating tasks. A recorded meeting goes in, a summary with action items comes out. Tools like Grain do this well and save real time.
All of them share a pattern: short scope, cheap errors, a person reviewing.
What does not yet
The agent that works alone for hours. The math is unforgiving: errors multiply step by step. A long chain without supervision ends up far from the goal, and the worst part is that it gets there with total confidence.
The agent with write permissions and no review. An agent that can send email, delete files or make charges needs human approval at the irreversible step. It is not distrust: the cost of an error is asymmetric.
Predictable cost. An agent that decides how many steps to take also decides how much to spend. Without hard limits configured, the bill is a surprise.
How to start well
Start with a normal flow that has one AI step, not with a full agent. In Activepieces or n8n that takes minutes to build and already gives you 80% of the value.
Once that flow works and you understand where it fails, hand the decision over to the model one step at a time.
Frequently asked questions
What exactly is an agent?
A model with access to tools and the ability to decide which one to use at each step until it meets a goal. It differs from a normal flow in that the sequence is not fixed in advance: the model decides it as it goes.
Where do they work today?
On narrow, repetitive tasks where an error has little cost: classifying, summarizing, extracting data, drafting reply drafts. Always with a person approving before the irreversible action.
And on long tasks?
That is where they fail. Each step introduces a probability of error, and that probability accumulates. Twenty steps with 95% accuracy each give less than 36% total accuracy.
How do I start without overspending?
With a spending limit on the API key and with the agent restricted to reading, not writing. When you trust what it proposes, you give it permission to act on one thing at a time.