Anthropic is changing the way its coding agent Claude Code works in two places at once. First, the so-called auto mode will become the default for the Pro, Max and Team plans from 14 August. Second, multiple sessions running in parallel can now communicate directly with each other. For teams that use AI agents in client projects, these are not cosmetic changes but interventions in approval and review processes.
Auto mode: a classifier takes over approvals
In auto mode, it is no longer the developer who decides at every step whether a command may be executed. Instead, a classifier — a model that sorts inputs into categories, here into non-critical and potentially dangerous commands — automatically checks the planned actions.
Anthropic explicitly cites safety as the reason for the move. In tests, according to the provider, the classifier recognised 89 per cent of dangerous commands, while human reviewers recognised only 13.6 per cent. The figures are remarkable because they reverse the widespread assumption that the human in the approval loop is automatically the safer instance.
This effect is plausible above all because of fatigue: anyone who clicks away confirmation dialogs every minute eventually stops reading them. Approval becomes a formality. The automated check is intended to replace precisely this routine.
In practice, this means that control shifts from the individual confirmation to the framework conditions. What becomes relevant is the environment in which the agent runs, what permissions it has there, which directories and credentials are reachable and what ultimately ends up in the repository. Anyone who wants to continue reading along step by step should check how the default setting can be adjusted in their own setup before the change takes effect.
Sessions that exchange information
The second innovation concerns the coordination of multiple agents. On macOS and Linux, Claude Code instances running in parallel can send each other messages, share findings and make status queries. Until now, such sessions worked largely in isolation; intermediate results had to be transferred manually.
This makes practical a pattern that many teams had already improvised anyway: one session analyses the existing code, a second builds the change, a third checks tests or documentation. The exchange now takes place within the tool instead of via copy-and-paste between terminal windows.
The downside is the reduced traceability. When sessions pass findings to one another, an internal flow of information arises that is no longer visible in the result. A false assumption from the first session can thus flow unnoticed into the work of the second. The information on platform support is clear: for Windows there is no statement on auto mode or session communication; macOS and Linux are named.
What the two announcements add up to
Both changes point in the same direction. The agent is to run independently for longer, be interrupted less and organise itself better internally. The human role thus shifts further from writing code to monitoring results — a development that is explicitly named in connection with auto mode.
The sources differ in level of detail. For auto mode, the date, affected plans and test results are documented. For session communication, only the function and the supported operating systems are available so far, with no information on limits, protocols or logging. Anyone who wants to use the function productively should clarify these points in their own test instead of assuming them.
Consequences for governance in client projects
When approvals are automated and agents coordinate with one another, control points must take effect elsewhere. The following are particularly sensible:
- Isolated execution environments with clearly limited permissions, separated from production systems and production credentials.
- Mandatory human code reviews before the merge, regardless of how many approvals the agent has granted itself.
- Automated safety nets in the pipeline: tests, static analysis, dependency and secret scans as a condition for acceptance.
- Documented responsibility: who is liable for changes that an agent introduced without individual confirmation?
- Traceability of which sessions were involved in a work result.
The question of client communication requires particular clarification. If an agent prepares changes in a project largely independently, the contract or project agreement should stipulate that AI-supported tools are used and what quality assurance looks like for them. This is less a technical question than one of trust.
Assessment for web and software projects
For TYPO3 and web projects, the change means that quality assurance shifts more into the pipeline and into review and less to the terminal. Anyone who still relies today on the person at the computer confirming every step loses this control point and should replace it with automated checks and a mandatory four-eyes principle before the merge.
Session communication is a productive lever for larger refactors, but it demands discipline in documentation so that it remains traceable why a change looks the way it does.