AI

AI agents are no longer just chatbots - they take on whole tasks, but someone has to keep an eye on them

Over the past two weeks, sources have described AI agents that handle customer calls, operate a computer and fix bugs on their own. The human role is shifting from doing the work to setting the rules and checking the result.

Bartosz Bielawski

Bartosz Bielawski

· 3 min read
Share
AI agents are no longer just chatbots - they take on whole tasks, but someone has to keep an eye on them

Signal in the data

Independent sources on this topic: 8 in the last week, 11 in 30 days, 20 in 90 days. We have tracked it since 2026-07-08; 52 claims in our database in total.

wk 310
wk 321
wk 334
wk 341
wk 352
wk 365
wk 374
wk 383
wk 390
wk 402
wk 413
wk 425

Number of independent sources per week (last 12 weeks).

What is happening

AI agents have been a talking point for months, but in October the tone changed. Instead of offering definitions, sources now show agents that complete whole tasks from start to finish.

One agent handles a customer call, records it and writes a summary (CallPage). Another opens apps, reads the screen, clicks and types on its own (Theo - t3.gg). Yet another fixed a bug in power plant software while working locally with no internet access (Adrian Andrzejczyk).

On top of that, OpenAI is releasing a drag-and-drop tool for building agents without writing code (Artur Jabłoński).

Where sources agree

Sources agree that an agent is not just a better chatbot but something that is given a task and plans the next steps itself. It is increasingly treated as an extra team member, or even a whole team, because it can work on several things at once (SemlyAI).

They also agree that the model is only one part of an agent. The rest is the tools around it, access to data and the rules it works by (Tech With Tim).

An agent that talks to customers also becomes a source of knowledge. It shows what customers ask about and what is missing from the product documentation (CallPage).

Where they differ

Sources differ on how much autonomy an agent should have. Some describe fully autonomous agents, for example ones that work on a website's visibility (SemlyAI).

Others point out that fully autonomous agents are still on the roadmap. Today, most systems require a human to review and approve each step (Goldie SEO Tips).

There is also a warning: an agent can take actions without the user knowing, which is a risk in itself (AI_Tech Leaders).

What to do about it

Start with one repetitive task that takes someone hours today, such as summarising customer calls. Describe exactly what a job done well looks like.

Give the agent clear limits: it answers only from your materials, asks when it is unsure and never makes things up. One tutorial demonstrates exactly these rules (Visual Studio Code).

At first, approve every action that reaches the outside world: a message to a customer, a change to the website, a payment. Give the agent more freedom only once its results are consistent.

Our take

AI agents are no longer a novelty and now take on whole tasks, from customer calls to operating a computer. The value comes not from the model itself but from a well-defined task and clear limits. The human does not disappear from the process - their role shifts from doing the work to setting the rules and signing off on the result.

Confidence: medium - there are many examples from different sources, but they are mostly demos and claims rather than measured results from companies. We will change our view if: data from real deployments showing that agents without human oversight perform as well as supervised ones, or that oversight eats up most of the time saved.

Sources

Read more