Tool Output Filtering and Node Approvals
Google released version 2.12.0 of its Agent Development Kit with mechanisms to govern tool interactions and workflow execution. Developers can now implement an after-tool callback to screen and modify tool outputs before returning the data to the model. This interception layer allows applications to scrub sensitive details or discard irrelevant content before the context enters the prompt history.
For workflow execution, ADK adds a confirmation flag supported across workflow nodes and FunctionNode instances. Teams can require explicit human confirmation prior to running a specific node. In parallel execution, the framework now pauses active workers when a human-in-the-loop interruption occurs, and a companion bug fix prevents confirmed tools from running more than once per granted approval.
Deployment Security and Session Latency
On the infrastructure side, the release permits teams to deploy agents directly to Agent Engine using dedicated custom service accounts, narrowing permissions instead of using broad default access. Security patches also address credential handling, including a fix that stops OAuth2 tokens from leaking across run and session endpoints.
Google reports that the new version reduces time to the first model request by 40 percent during long sessions. The runtime environment also expands with verified support for installing and running on Python 3.15.
Our assessment
Teams operating long-running agent workflows should test whether output filtering alters context quality or breaks downstream assumptions in complex prompts. Operators deploying to production will need to verify whether existing human-in-the-loop approval handlers properly capture node pauses under parallel execution. Finally, systems that previously relied on ambient cloud credentials should migrate explicitly to custom service accounts to limit execution privileges.
Source: Google original publication.
Read the original material
Google (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.
