Domo’s data team adds shared folder, naming, test and deployment conventions above permission-aware MCP tools—separating access control from maintainability.

Databricks’ five-signal scorecard is a better starting point for measuring whether natural-language analytics changes decisions without multiplying metric drift.

AI.IF can route most rows through a distilled local model, but that default-looking optimization changes scale requirements, error handling and endpoint choices.

OpenSearch 3.7 puts prompts, tool calls, span links and logs in one analytical store. AWS’s own engine-selection guidance draws a firm boundary around where it fits.

A Saturday checkpoint found no new product delta across AWS, Databricks, Snowflake, Oracle and IBM—but the feeds’ different clocks make “nothing shipped” a narrower claim than it sounds.

The same question-and-query pair can act as a retrieved hint, a composable blueprint or a governed semantic-model asset. Treating those roles as interchangeable is a migration trap.

Databricks, Confluent and Google all promise current data without traditional batch copies—but they differ on checkpoints, maintained state, replay and who owns correctness.

Feature Views reduce training-serving skew by reusing definitions across historical and online paths. They do not make asynchronous feature delivery disappear.

A new impossibility result isolates bag difference as the boundary: canonical and compatible interpretation semantics cannot recover SQL’s monus operation.

Three launches from Microsoft, Google and AWS expose the same deployment lesson: moving data or metadata is not the same as moving its original access rules.

The Slack and Teams previews evaluate questions with the asker’s Unity Catalog rights, then may show the resulting answer to people who could not query the source themselves.

The proposal separates continuously maintained business state from transaction authority. That distinction matters more than the new category name.

The new Apache-2.0 project treats semantic definitions and read-only execution policy as infrastructure, not prompt context. Its repository is substantive, but release and adoption evidence are absent.

A two-tool GraphRAG design outscored a seven-tool knowledge-graph agent in Snowflake’s controlled biomedical test. The useful lesson is architectural, not a general leaderboard claim.

Fabric can now share an entire lakehouse or warehouse without copying it. The preview’s operating boundary is equally important: source OneLake rules do not follow the data, and agents need an extra shortcut layer.

The launch separates a fast decision model from a general-purpose LLM, but operators should not treat the cited 26 ms routing figure as a measured Koa production SLA.

Across six 80-case evaluation suites, a learned router spent fewer tokens than always-on retrieval; the result is a warning against treating tool access as automatically useful.

Version 5.3.9 beta 2 adds checkpoints, rejected-row exports and clearer counters. The useful change is not “retry”; it is making partial success inspectable.

The preview replaces custom MCP plumbing with first-party live-query connectors. Its convenience creates a deliberate boundary: broad access follows the user, while deterministic scope requires a curated data agent.

The same Claude model improved sharply with governed context, but the vendor’s filtered four-domain BIRD experiment should guide architecture decisions, not cross-system rankings.

Four enterprise implementations converge on the same design: business meaning must be machine-readable, queryable and tied to security before an agent can act safely.

TSIQ’s two-plane design is a useful production pattern: let models interpret requests, but keep market data, calculations, permissions and audit trails deterministic.

The unified engine matched Chroma’s vector recall and beat LadybugDB on transactional graph work, but specialist systems remained dramatically faster on analytics.

MSK Replicator now spans both AWS GovCloud regions and carries offsets, ACLs and topic settings. Teams still need to design around asynchronous replication and a documented new-topic window.

The headline latency joins generally available ingestion and serving to a transformation layer that remains in private preview. Teams should qualify each stage separately.

Tableau Knowledge can ground external agents in governed analytics assets, yet its deployment, licensing and connector boundaries matter before teams treat it as a universal semantic layer.

Databricks’ new OpenSharing walkthrough makes one operational boundary explicit: incremental append processing, row-level CDC and live querying are different contracts.

For schema-heavy data agents, the headline context window is only part of the deployment contract: explicit caching skips Converse, reasoning defaults to maximum, and every supported route is cross-Region.

PQR-Search’s authors replace average ANN recall with a per-query probability contract, then show what that guarantee costs across static and streaming workloads.

The new limit is opt-in, provisioned clusters must reboot, and VARCHAR(MAX) still means two different things depending on context.

Agent Registry is now in GovCloud US-West. The useful boundary is also the easy one to miss: catalog approval governs discoverability, while execution still belongs to separate runtime, gateway and identity controls.

Unity Gateway centralizes model access, budgets and traces; public-sector teams still need a route-by-route residency inventory.

Cloud 26.10 makes custom Spotter Analysts generally available while adding user- and group-level connector sharing. Those controls solve different problems, and teams should test both.

The new API compresses forecasting, prediction and natural-language filtering into familiar pandas calls. Teams still need to govern where the data goes, which model runs and what the result means.

Source-vended credentials remove the need for a broad AWS role over remote storage. They also make operators reason about revocation, query-time refresh and layered grants across two systems.

CoRe’s controlled renaming test shows that some LLM repair scores depend on familiar entities even when the candidate values remain in the table.

IBM’s Think 2026 data recap makes a useful operational point: freshness, context and governance are separate controls, and autonomous action needs all three.

The preview combines plans, schema and runtime evidence with an LLM, but it does not turn recommendations into verified changes.

Microsoft’s preview exposes timeout, retry and polling controls. Those controls improve delivery; they do not prove an external write is idempotent.

The young LangGraph project is notable less for its model layer than for a typed evidence contract, explicit execution boundaries and unusually candid validation limits.

The reported accuracy gap is real in the test, but cleaner data modeling explains much of the improvement before MetricFlow enters the path.

The new BigQuery endpoint covers agent resources at rest and jurisdictional model processing in use. It does not relocate datasets or extend the same commitment to Looker.

A 454-case held-out test separates diagnosing a database fault from actually fixing it, but its composite score and LLM judge need to travel with every comparison.

Google’s open-source agent analytics stack turns requests, tool calls and errors into queryable data. Teams still have to decide what success means—and who may inspect the logs.

The Dedicated Exadata feature dynamically chooses specialist workers, but every handoff is sequential and each worker receives only the context the supervisor packages.

Oracle, Databricks and Snowflake expose three different context mechanisms. Treating them as substitutes leaves predictable gaps in names, joins and answer patterns.

A new semantic-filtering study reports the best accuracy in 12 of 13 settings, but its own audit says one setting missed the oracle-relative target in 30% of repeated draws.

NVIDIA cut a protected 1.2.0 branch and synchronized its app, APIs, MCP package and Helm images. Operators still need to verify the artifact they deploy.

OpenSharing, Clean Rooms and Marketplace can all move data and AI assets, but they expose very different things to very different counterparties.

The private-region launch moves compute, storage and encryption keys into a customer account while leaving identity, billing, telemetry and optional AI processing across the control-plane boundary.