Original news analysis | Symaxx Digital | 8 November 2026
Google’s latest Python agent toolkit adds a Microsoft Teams connection and tools for testing the same task repeatedly. For South African businesses, the useful question is whether a small internal workflow can become more dependable without creating an expensive support problem.
Google posted the ADK Python 2.14.0 release on GitHub on 7 November. Its changes include TeamsRunner, multi-run evaluations and more flexible checks on the arguments an agent passes to its tools. The release’s own changelog heading is dated 4 November. Google’s release notes
ADK is an open-source development framework for building and running custom agents. This is a developer release, so putting it to work still requires an implemented application. We have reviewed the documentation, but have not tested this version in a South African business or Microsoft 365 tenant. ADK overview
What changes for a business team
TeamsRunner lets people speak to an ADK agent in personal chats, group chats and channels. Google’s setup uses the Microsoft 365 Agents SDK, a registered bot and an HTTPS endpoint hosted by the business or its provider. Teams integration guide
That creates a possible route for an internal assistant that answers questions about approved procedures or helps staff find the right service information. The business should decide which information it can access and what it may do before connecting operational systems.
The evaluation update is particularly useful for that decision. Teams can run each test case several times and combine scores or pass rates, rather than judging an agent from one successful demonstration. Evaluation changes in 2.14.0
Google also documents argument checks that distinguish exact matches, numeric tolerances and meaning-based comparisons. A model can judge whether two differently worded search queries mean the same thing. For a pilot, we would keep customer identifiers, prices and permission boundaries exact, even when allowing flexibility in ordinary wording. Tool argument matching
Teams support has practical limits
The documented Teams connection exchanges text only. Incoming attachments, images and Adaptive Card submissions are not passed to the agent, and non-text agent output is not posted back. A team that mainly works through uploaded quotes or screenshots should check this before choosing its first use case.
Hosting also matters. Google warns that an in-progress reply is lost if its server process stops. Platforms that suspend background processing can interrupt the work. In-memory conversation history also disappears on restart and is not shared across replicas; Google recommends persistent session storage for multiple instances. Teams limitations
Ask the implementer to show what staff see after a restart or failed response. A polished chat window does not establish that a request completed correctly.
South African access and the real cost
Google’s indexed official Gemini API availability page includes South Africa. That supports Gemini as a possible model route, but does not verify access to every model, hosting region or company tenant. The live availability page could not be rendered during this review, so confirm current account eligibility before committing. Google’s supported-region list
ADK’s Python code uses the Apache 2.0 licence. There is no separate ADK software licence charge, but that does not cover the model, infrastructure or implementation. Google publishes Gemini API usage prices in US dollars, with charges depending on the selected model and services. ADK repository and licence · Gemini API pricing
Request a budget that separates development, model usage, hosting, storage, Microsoft account requirements and ongoing support. Include evaluation runs in the usage estimate. For a rand budget, state the exchange-rate assumption and applicable taxes. There is no credible fixed monthly total until the workflow, model and expected volume are defined.
A sensible first pilot
We would start with one internal, text-based task and a named business owner. For example, an assistant could answer questions from an approved service handbook, with a staff member responsible for corrections.
Before launch:
Build a test set from realistic questions, including ambiguous requests, missing information and contradictory instructions.
Repeat those tests and record both correct answers and failures. Review the underlying responses, not only the combined score.
Include questions the assistant should decline or refer to a person. Keep record changes and external messages outside the first pilot.
Check that one employee cannot obtain information they should not see, even when asking indirectly.
Test restarts, slow responses and duplicate requests. Give staff a clear route to report a failed or misleading answer.
Compare useful answers, review time and operating cost with the existing process before expanding access.
These are proposed acceptance checks, not results we have measured with ADK 2.14.0.
Existing ADK users should review the upgrade
The release includes breaking changes. Developers should check tools that expect string values from enum arguments and applications using Vertex AI session configuration. Google also warns that Firestore sessions written by this version cannot be read by version 2.13.0 or earlier after a rollback. A rollback plan therefore needs to cover stored data as well as application code. Breaking and behaviour changes
For a South African team already using Teams, this is a useful development option to assess. The strongest starting point is a narrow service that staff can test repeatedly, with clear ownership and a budget for keeping it reliable.