Google AI Updates: November 7, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Google Finds Stale TODO Comments Send Its Coding Agents on Detours in 6% of Exposed Sessions, Costing 77% More Tokens
Google engineers analyzed distributed traces from thousands of internal engineers and hundreds of millions of agent steps to measure how inline TODO(bug/...) and FIXME(bug/...) references distract coding agents. In 13% of file-reading sessions the agent saw such a reference; it ignored the comment 94% of the time, but in the other 6% it took action on that bug ID, and 83% of those actions were unprompted by the developer. Of the unprompted detours, 41% chased tickets that were already closed (32% fixed or verified, 9% obsolete or won’t-fix) and 57% involved open bugs where the agent widened its scope, looking them up in 61% of cases and editing code in 39%. A fixed-effects regression controlling for developer, codebase complexity, task type, model tier, files read, prompt length and step position found these detours add 70% more steps, 77% more tokens and 105% more wait time per session (p<0.0001). Google responded with cleanup agents that check bug, file and change history before proposing deletions, which removed thousands of obsolete references over six months with more than 95% reviewer acceptance. Source
2. Google Compute Engine’s Z4M Reaches GA With 168,000 GiB of Local SSD and 400 Gbps RDMA for AI Checkpointing and KV Cache
Google Cloud made the storage-optimized Z4M machine series generally available as VMs and bare-metal instances, aimed at AI training and inference storage such as checkpointing, distributed KV cache and parallel file systems. Built on fifth-generation Intel Xeon Scalable processors and Titanium SSDs, Z4M offers up to 192 vCPUs, 1,536 GiB of memory and 168,000 GiB of Local SSD (875 GiB per vCPU), with up to 15.6M random-read IOPS and 75,600 MiB/s sequential reads, which Google puts at up to 70% and 110% above Z3. It pairs that storage with up to 400 Gbps of Cloud RDMA networking, which Google says cuts data-access latency by up to 60% versus TCP, and supports NVIDIA GPUDirect Storage and TPUDirect Storage to keep accelerators fed. Google also positions the bare-metal shapes for running isolated agent sandboxes at scale, including with Agent Substrate. Source
3. Gemini for Business Previews Data Cloud Connectors That Run Permission-Scoped SQL on BigQuery, Spanner, AlloyDB, Cloud SQL and Looker
Google Cloud put Data Cloud connectors in Gemini for business into preview, letting users ask natural-language questions of live enterprise data from Gemini without knowing which table or project holds it. Gemini searches Knowledge Catalog to find candidate data, maps everyday terms onto schema using technical metadata, business glossaries, semantic models and verified queries, then generates SQL and runs it through the connector under the end user’s own credentials, so existing row- and column-level permissions apply. Sources include BigQuery, Spanner, AlloyDB, Cloud SQL and governed Looker data (several in preview), and results can be pushed into a Google Doc, a Gmail draft or a Slides deck. The connectors sit alongside Conversational Analytics agents, which carry use-case-specific context and instructions. Source
4. BigQuery’s AI.AGG Preview Summarizes Millions of Rows of Text or Images in One SQL Query Through Recursive Gemini Batching
Google Cloud detailed AI.AGG, a preview GoogleSQL aggregate function that applies one natural-language instruction across millions of rows of unstructured or multimodal data, which would not fit in a single model context window. BigQuery rejects or drops rows over 10 MiB, groups the rest into batches sized to the chosen Gemini endpoint’s token limits, runs the prompt on each batch in parallel through Vertex AI, then feeds the intermediate summaries back to the model with the same instruction until one aggregated string remains per GROUP BY group. Unlike row-level functions such as AI.CLASSIFY, which can use distilled CPU proxy models, AI.AGG uses full-weight LLMs at every stage, and it reads images and other files through Object Tables and OBJ.GET_ACCESS_URL. The design replaces the external map-reduce pipelines, rate-limit backoff and chunk stitching teams previously built themselves. Source
5. BigQuery DataFrames Adds a .bigquery Pandas Accessor That Runs Zero-Shot TabFM Predictions in One Call
Google Cloud added a .bigquery accessor to BigQuery DataFrames (BigFrames) so pandas users can call BigQuery AI functions directly on ordinary DataFrames. Importing bigframes.pandas and setting a project enables calls such as train_df.bigquery.ai.predict(new_data_df, label_col='target_metric'), which replaces a local scikit-learn training pipeline with a zero-shot prediction from Google’s TabFM tabular foundation model. The accessor translates the operation into GoogleSQL, runs the model on BigQuery and returns a standard pandas DataFrame, and Google says it also covers forecasting, classification and embedding generation. Source
6. Google’s ADK for Python 2.14 Graduates Sandboxed Execution Environments, A2A and Evaluation From Experimental
Google published Agent Development Kit for Python v2.14.0 on GitHub on November 7, moving 26 auth, environment, app, CLI and cache modules, 16 toolset and integration features, and the evaluation and user-simulation classes out of experimental status. The execution-environment layer (BaseEnvironment, LocalEnvironment, EnvironmentToolset, DaytonaEnvironment, E2BEnvironment) is now stable and streams file reads in chunks, the A2A integration is stable with an adk generate_agent_card command, and evaluation gains adk eval --num_runs with --parallelism to run each case several times and combine scores, plus a ToolArgsMatcher that compares tool-call arguments exactly, case-insensitively, with numeric tolerance or through an LLM-backed semantic check. New integrations include a TeamsRunner for Microsoft Teams, an OpenAIDecisionsClassifier for workflow and tool-gating questions, and vector retrieval with TTL expiry in MongoDbMemoryService, and Google reports a 200-turn run of a 30-node workflow is 23% faster with 9% lower peak memory. The release has breaking changes: FunctionTool parameters typed as Enum now receive enum members instead of raw strings, VertexAiSessionService.create_session rejects extra config keys such as http_options, and RunSkillScriptTool in environment mode takes structured arguments instead of a free-form shell command. Source