Within a few days, two manufacturers have moved their development tools in the same direction: AI agents no longer just demonstrate tasks, but complete them independently – in the terminal, in version control and in the browser. What previously ended up as a suggestion in the chat window is now executed.

Claude Code asks less often

Anthropic enables the so-called Auto Mode in Claude Code by default. Claude Code is the provider's command-line-based development tool; Auto Mode means that the AI independently executes shell commands, Git operations and tool calls instead of obtaining confirmation for every step.

The practical effect is obvious: fewer queries, fewer interruptions, longer coherent work steps. The price is equally obvious. Anyone who delegates shell and Git access delegates the ability to modify files, install dependencies and create commits. Control thus shifts from individual approval to prior configuration and subsequent review.

WebStorm agents debug in Chrome

In parallel, JetBrains is expanding the JavaScript IDE WebStorm. Agents can now be controlled there via specifications – that is, via a written description of what is to be achieved, instead of individual prompts in the dialogue. For debugging, the agents can also access the Chrome browser directly.

This closes a gap that has been particularly noticeable in web development: an agent that only sees the source code can merely guess at runtime errors in the browser. An agent that can actually load the page, observe the behaviour and read error messages works with the same information as a developer in front of the same screen.

The specification-driven approach deserves particular attention. It shifts effort forward: anyone who describes precisely what should apply obtains more traceable results than with a chain of improvised prompts. For projects with multiple participants, this is an advantage because the specification can be documented and reviewed – the chat history, by contrast, hardly.

The browser becomes the workspace

The third announcement shows that this development is not limited to development environments. Anthropic's Claude Cowork now also runs in the side panel of the company's own Chrome extension. The agent thus sits directly next to the application with which work is being done, instead of in a separate window.

The three announcements concern different products and should not be read as one package. What they share is the direction: agents are moving to where the work actually takes place – into the terminal, into the IDE, into the browser – and they act there with fewer intermediate approvals than before.

What this means for ongoing projects

For web projects, the benefit is concrete. Reproducible frontend errors, faulty network calls, regressions after an update: tasks whose diagnosis previously required a lot of manual clicking and checking can be delegated. The time from the error message to a reliable root cause analysis can thereby drop significantly.

At the same time, the risk question shifts. It is no longer whether an agent can technically solve a task, but what it is permitted to do so with. Four points are decisive for this in practice:

  • Separate permissions: Agents work on local working copies and development environments, not with credentials for production systems or live databases. What an agent cannot reach, it cannot damage.
  • Use Git as a safety net: Separate branches, no direct pushes to main branches, protection rules in the repository. Automatically generated changes should be as traceable as any other change – ideally recognisable as such.
  • Do not give up review: Autonomous execution does not replace professional acceptance. Precisely because more code changes arise in a shorter time, review becomes the bottleneck. Anyone who saves here only shifts errors to later.
  • Assess data access in the browser: Agents in the browser panel potentially see what is open in the browser. For environments with customer data, personnel data or contract documents, a deliberate decision is needed about where such extensions may be active – and where not.

There is also an organisational point: when a tool changes its default setting, behaviour changes in existing setups without anyone actively having changed anything. Default values therefore belong in the internal documentation and on the test bench during updates. It is worth recording in the team which modes are permitted in which projects, instead of leaving that to each individual workstation.

Keep expectations realistic

Autonomy does not mean reliability. An agent that acts without asking can execute a chain of plausible but incorrect steps – and faster than before. Experience shows that a staged approach therefore proves effective: first in clearly delimited, well-tested areas, then on a larger scale. Automated tests and a functioning continuous integration pipeline, i.e. the automatic checking of every change, are less of an extra than a prerequisite.

Equally important is managing expectations towards clients. Faster error fixing is a real promise.

But it does not arise from eliminating review steps; it arises from eliminating mechanical work and concentrating the remaining attention on evaluation and decision-making.

For web and software projects, this development means above all one thing: the bottleneck shifts from writing to reviewing and approving. Anyone who now clarifies permission concepts, branch strategies and review rules can take advantage of the speed gain without losing traceability. Anyone who adopts the defaults without examining them brings both speed and ambiguity into the project at the same time.

Sources