AI

AI agents have stopped being chatbots - now they work in your company, your browser and your reports

17 agents at Levi Strauss cut a project from 9 months to 2.5, and about half of Notion pages are now created by agents. Several independent sources describe the same shift this week: the agent is leaving the chat window and becoming a task executor with a defined input and output.

Bartosz Bielawski

Bartosz Bielawski

· 3 min read
Share
AI agents have stopped being chatbots - now they work in your company, your browser and your reports

Signal in the data

Independent sources on this topic: 5 in the last week, 12 in 30 days, 20 in 90 days. We have tracked it since 2026-07-08; 34 claims in our database in total.

wk 340
wk 351
wk 363
wk 371
wk 381
wk 390
wk 401
wk 411
wk 427
wk 434
wk 444
wk 451

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

What is happening

A year ago, an AI agent meant a chat window where you had to sit and correct the answers. Now the sources describe something different: the agent gets a task, does the work and hands the result to someone who only has to approve it.

The clearest example comes from Levi Strauss. The company built 17 specialised agents covering everything from requirements through testing and accessibility to security. Throughput rose 51%, and the share of AI-assisted changes went from 26% to 69%. The new version of an app took 2.5 months, compared with 9 months for the old one.

The same shift shows up in everyday products. Notion announced a new chapter and said AI usage inside the app had grown ninefold, with roughly half of pages now created by agents. Cloudflare, meanwhile, is already charging AI companies for signals about when a page has changed, so that an agent does not fetch it again needlessly.

Where the sources agree

First: an agent will not fix a mess. Several authors independently make the same point - the quality of the output depends on the quality of the data and the process, not on the model. An agent completes the task exactly as written and does not aim for a better result unless someone asks for one.

Second: an agent needs boundaries, just like an employee. The same theme keeps coming up - permissions, hard limits and a clearly defined scope are a precondition for deployment, not an obstacle to it.

Third: humans stay where the decisions are made. At Levi Strauss, the whole process was reduced to two human moments: approving the specification and the final merge. Everything else is handled by the agents, but accountability does not change hands.

Where they disagree

The sources disagree on where to deploy agents first. Some point to online shopping, travel and knowledge work; others to boring, repetitive tasks that free people up for creative work. For a service business, those point in different directions.

The second dispute is about control. Some authors trust a repository and reviewed changes, because versioning and auditing are familiar. Others say plainly that current developer tools were not built for many agents working in parallel and that new ones are needed.

The third dispute is the sharpest: autonomy versus security. A browser agent can be attacked through a hidden instruction on a page it reads. The more you allow it to do, the larger the attack surface - and no source claims this problem has been solved.

What to do about it

Start with one task that repeats every week and describe it as a process: what triggers it, what it receives, what it should do, what it hands back and to whom. Only then look for a tool.

Decide where a human makes the decisions and where a human only carries out the work. If you do not separate the two, the agent will end up making decisions nobody meant to hand over.

Finally, measure it the same way you measure people: time, quality and the number of corrections. A deployment without measurement only looks good in a presentation.

Our take

Agents are leaving the chat and entering processes, but the value comes from the work described around them rather than from the model: the trigger, the input, the output and the point where a human makes the decision. Companies that start with the process rather than the tool see a measurable effect within weeks. Those that start by buying a licence get a demo.

Confidence: medium - the direction is confirmed by several independent sources and one concrete deployment, but most of the numbers come from tool vendors. We will change our view if: Agent deployments in service businesses (not software) that show no impact on time or quality after a quarter would disprove the shift from chat to executor.

Sources

Read more