live wire
IBM makes watsonx Orchestrate AgentOps, custom LLM judging and Bedrock-agent discovery generally availableIBMSchemaGate 0.1.45 fixes broken Oracle ADB wallet connections and an OCI stack pinned 28 releases behindSchemaGatePDI’s Amazon Quick procurement agent grounds spend answers in vendor, category and contract contextAWS Business Intelligence BlogBigQuery’s ML.METRICS example returns 0.84 accuracy but 0.30 macro-F1 on the same 100-row classification queryGoogle Cloud BigQuery docsSchemaGate 0.1.44 auto-selects sentence embeddings, lifting bundled-schema retrieval from 90/98 to 93/98SchemaGateSchemaGate 0.1.43 adds read-only SQL execution with per-principal table checks—and documents unauthenticated client assertionsSchemaGateDatabox adds reusable AI Analyst Skills with personal/company scope, auto-matching and marketplace installsDataboxFabric previews an AI builder for data-agent instructions, source guidance and example queriesMicrosoft FabricDatabricks trains data-agent retriever to stop early or spend bounded extra search steps, reporting 5.8-second latencyDatabricksThoughtSpot adds SpotterCode coding agent to its Visual Embed PlaygroundThoughtSpotLongMemEval-S audit: 67–73% of restore-fixable 80k-budget errors came from evicted evidence under three policiesarXivSchemaGate 0.1.42 adds dimension-aware retrieval and fixes complex multi-table SQL promptsSchemaGateSnowflake agent toolsets can silently drop inherited tools when callers lack accessSnowflake DocumentationLooker’s VS Code extension reaches GA with MCP-assisted LookML generation, editing and validationGoogle Cloud Looker release docsIBM makes watsonx Orchestrate AgentOps, custom LLM judging and Bedrock-agent discovery generally availableIBMSchemaGate 0.1.45 fixes broken Oracle ADB wallet connections and an OCI stack pinned 28 releases behindSchemaGatePDI’s Amazon Quick procurement agent grounds spend answers in vendor, category and contract contextAWS Business Intelligence BlogBigQuery’s ML.METRICS example returns 0.84 accuracy but 0.30 macro-F1 on the same 100-row classification queryGoogle Cloud BigQuery docsSchemaGate 0.1.44 auto-selects sentence embeddings, lifting bundled-schema retrieval from 90/98 to 93/98SchemaGateSchemaGate 0.1.43 adds read-only SQL execution with per-principal table checks—and documents unauthenticated client assertionsSchemaGateDatabox adds reusable AI Analyst Skills with personal/company scope, auto-matching and marketplace installsDataboxFabric previews an AI builder for data-agent instructions, source guidance and example queriesMicrosoft FabricDatabricks trains data-agent retriever to stop early or spend bounded extra search steps, reporting 5.8-second latencyDatabricksThoughtSpot adds SpotterCode coding agent to its Visual Embed PlaygroundThoughtSpotLongMemEval-S audit: 67–73% of restore-fixable 80k-budget errors came from evicted evidence under three policiesarXivSchemaGate 0.1.42 adds dimension-aware retrieval and fixes complex multi-table SQL promptsSchemaGateSnowflake agent toolsets can silently drop inherited tools when callers lack accessSnowflake DocumentationLooker’s VS Code extension reaches GA with MCP-assisted LookML generation, editing and validationGoogle Cloud Looker release docs
nl2sql.ai
releaseRelease analysis

Snowflake makes partial answers the default when Cortex Agents lose a tool

The September 2 behavior change preserves role-based access, but moves a missing grant from an HTTP failure into the answer stream — so clients must treat warnings as part of correctness.

Partial answers replace hard failures when a tool is inaccessible.
AI-generated illustration
By The News Desk· Sep 4, 2026the quick take — two AI hosts go live when you do

Snowflake is changing how Cortex Agents and Snowflake CoWork behave when a caller cannot access every tool configured on an agent. Starting September 2, 2026, the default is to continue with the tools the caller’s role can use, return an answer, and emit a warning for each inaccessible tool. Snowflake says the change applies to agent runs through REST, SQL and CoWork conversations. Snowflake’s behavior-change notice

From hard failure to degraded execution

Previously, a missing privilege on a configured Cortex Search, Cortex Analyst, MCP or skill tool caused the request to stop at the first inaccessible tool with an HTTP 4XX response. Under the new default, the run returns HTTP 200 and proceeds with a smaller tool set. Streaming clients receive a response.warning event with code 399569; final REST responses and SNOWFLAKE.CORTEX.DATA_AGENT_RUN JSON place the warning in a top-level warnings array. Snowflake documents the before-and-after behavior

The security boundary does not change: Snowflake says no privilege is added, and callers still need the same grants to use each tool. What changes is the failure contract. One agent can now serve roles with different privileges without requiring a separate agent for each privilege class, but an apparently successful response may have been produced without a data source or analytical tool that the developer expected. Snowflake’s stated rationale and impact

Warnings become part of answer quality

For text-to-SQL systems, that makes warning handling a correctness concern rather than just an observability detail. A client that previously surfaced missing grants by catching HTTP errors will no longer see that signal. Snowflake explicitly tells developers to subscribe to response.warning and handle code 399569; the warning is also written to SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS for later audit. Snowflake’s client guidance

Teams that prefer fail-closed behavior can set orchestration.tool_not_accessible to reject, which returns one error listing every inaccessible checked tool, or to legacy, which preserves first-failure behavior. The new accept mode is the default when the field is unset. Snowflake notes that committed agent versions are immutable, so changing a versioned agent requires creating and committing a new version and repointing its default version or alias. Configuration and migration steps

The practical migration test is straightforward: inventory clients that equate HTTP 200 with a complete tool run, exercise each supported caller role against the agent specification, and decide whether a partial answer is safe for that use case. For informational assistants, degradation may be useful. For regulated or decision-driving analytics, reject may remain the clearer contract until the application can display and act on tool-access warnings.

Filed by The News Desk. Corrections: desk@nl2sql.ai · Our standards →

comments · 0

    Comments are moderated before they appear. Your email is used once to confirm it is you — never shown, never sold. Corrections and questions get an answer from the desk when we have one.