Can your AI governance stop an action, or can it only prove one happened?

In the first week of July, the two most widely deployed coding agents both shipped governance updates.

Anthropic moved Claude Code to manual approval by default. A human authorises file writes, shell commands and network calls before they run. The change followed reporting that the automatic safety classifier had missed 17% of overeager actions.

GitHub shipped Copilot agent session streaming. Enterprises can now pull every prompt, response and tool call into their SIEM. Each agent action, recorded and retrievable.

Both are real improvements. Both are needed. But they are not the same kind of control.

One works before the action. It can halt a command that should not run.

The other works after the action. It can prove the command ran.

Most of what gets called AI governance today is the second kind. Session logs and dashboards have value, but none of it stops anything.

Here is the test worth applying to any control in your pipeline. Can it halt the action, or can it only show the action was written down? If every answer is the second one, you have a record of what your agents did. You do not have a say in what they do next.

The most telling signal is Anthropic’s. When the vendor building one of the most capable agents decides the default should be a human check before execution, that tells you where they believe the risk sits. Not in the model. At the point of action.

Where in your pipeline is the check that can actually say no?