Google’s Python Agent Development Kit 2.11.0 adds workflow controls and memory storage. We have not tested these changes.
GitHub’s publication date is 2 October; the changelog heading says 1 October. This brief uses GitHub’s date.
Stop a run or pause for a decision
An abort_signal supports cancellation in runners, workflows and nodes. The /run_sse endpoint cancels work when its client disconnects.
Tool nodes can pause for approval through RequestInput, matching the existing confirmation behaviour of LlmAgent.
Consultation budgets and stored memory
ModelConsultTool supports asking another model for help, with turn and session limits. A sqlite:// URI selects the new local database memory service.
Google also lists optional MCP SDK 2.x connections.
Our assessment
For your next workflow trial, test the decisions around a tool call as carefully as its successful result. Use a harmless action, such as writing to a test document. Check what happens when you approve, refuse, delay your response or close the client. Record which steps complete and which remain pending.
Do not assume a cancellation signal reverses an action that has already reached another service. Your application may need to check that service’s state and provide a separate recovery step. Include that situation in your tests before allowing changes to important records.
For model consultation, define what information may be sent and set a small initial call budget. Check whether the additional response improves the task enough to justify its time and cost. Request limits provide a way to control calls; they do not demonstrate better answers.
For stored memory, decide what should survive between sessions and who can read or delete it. Try a restart and a deletion exercise with test data. Add these checks to your operating procedure so stored context remains an intentional part of the application.
Keep a record of the choices you make in that trial: permitted actions, approval points, recovery steps and memory retention. Repeat the same checks after an upgrade. Assign someone to check the results before rollout. This is a suggested evaluation plan; it does not establish how reliably any particular agent will behave.
Source: Google’s ADK Python 2.11.0 release notes.
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.
