live wire
▸Snowflake agent evals can pin judge versions; unpinned metrics may roll forward and shift scoresSnowflake▸ThoughtSpot makes Spotter 3 default for new instances; it requires download rights and always shares query data with an LLMThoughtSpot documentation▸MotherDuck adds image classification and object location to prompt_jev(), returning typed SQL outputs with confidence scoresMotherDuck▸BigQuery conversational analytics adds Montréal and Toronto; mixed-region agents still default to globalGoogle Cloud Documentation▸Fabric enables Azure OpenAI Copilot access by default, but OpenAI-subprocessor and cross-region processing/storage switches remain offMicrosoft Learn▸Snowflake Cortex Agents now continue by default when named tools are inaccessible, emitting warnings instead of failing the runSnowflake documentation▸Databricks Genie One can use group memberships to tailor answers, while Unity Catalog remains the access-control boundaryDatabricks documentation▸Fabric data agents can query Power BI models with Read alone; Build and workspace access are unnecessaryMicrosoft Learn▸Fabric data-agent Git configs expose column flags, but selecting a table still includes every columnMicrosoft Learn▸Google Data Agent Kit offers remote Spark kernels, but notebook exports can still move results into local Pandas or ArrowGoogle Cloud Documentation▸Cube says production analytics agents need explicit partial-result states, caller-filtered model search and durable artifactsCube▸Thinking Machines names Bridgewater behind ReViSQL-K2.6, citing 92.97% SC-16 accuracy at 15% of frontier-model costThinking Machines Lab▸AWS says a mismatched data/tool/prompt release cut its Field Sales agent pass rate from 78% to 38% in pre-productionAWS Artificial Intelligence Blog▸NL2SQL Multi-Agent ships evaluation artifacts across BIRD, CSpider, Archer and KDD DataSpace—but no software licenseGitHub▸Snowflake agent evals can pin judge versions; unpinned metrics may roll forward and shift scoresSnowflake▸ThoughtSpot makes Spotter 3 default for new instances; it requires download rights and always shares query data with an LLMThoughtSpot documentation▸MotherDuck adds image classification and object location to prompt_jev(), returning typed SQL outputs with confidence scoresMotherDuck▸BigQuery conversational analytics adds Montréal and Toronto; mixed-region agents still default to globalGoogle Cloud Documentation▸Fabric enables Azure OpenAI Copilot access by default, but OpenAI-subprocessor and cross-region processing/storage switches remain offMicrosoft Learn▸Snowflake Cortex Agents now continue by default when named tools are inaccessible, emitting warnings instead of failing the runSnowflake documentation▸Databricks Genie One can use group memberships to tailor answers, while Unity Catalog remains the access-control boundaryDatabricks documentation▸Fabric data agents can query Power BI models with Read alone; Build and workspace access are unnecessaryMicrosoft Learn▸Fabric data-agent Git configs expose column flags, but selecting a table still includes every columnMicrosoft Learn▸Google Data Agent Kit offers remote Spark kernels, but notebook exports can still move results into local Pandas or ArrowGoogle Cloud Documentation▸Cube says production analytics agents need explicit partial-result states, caller-filtered model search and durable artifactsCube▸Thinking Machines names Bridgewater behind ReViSQL-K2.6, citing 92.97% SC-16 accuracy at 15% of frontier-model costThinking Machines Lab▸AWS says a mismatched data/tool/prompt release cut its Field Sales agent pass rate from 78% to 38% in pre-productionAWS Artificial Intelligence Blog▸NL2SQL Multi-Agent ships evaluation artifacts across BIRD, CSpider, Archer and KDD DataSpace—but no software licenseGitHub
nl2sql.ai
releaseRelease

TrueForge 0.2.0 turns an open agent harness into an operations-ready scheduler

The stable release adds persisted cron jobs, ownership-aware authorization and metrics—the controls that matter after an analytics agent leaves the demo.

Before-and-after of an agent harness becoming production-ready.
AI-generated illustration
By The News Desk· Sep 20, 2026the quick take — two AI hosts go live when you do

TrueFoundry’s open-source agent harness reached a stable 0.2.0 release on September 18 with a decidedly operational feature set: persisted cron schedules, session metrics, tighter resource ownership and a Helm chart promoted from release candidate to stable. The release notes describe a production-oriented step rather than a new model or prompting technique.

Scheduling becomes a first-class object

TrueForge can now save cron schedules for existing agents, expose them through /api/v1/schedules, validate scheduling policy at write time and claim due runs through a single dispatcher. In distributed deployments, a dedicated controller process runs those periodic loops while the server handles requests; standalone deployments keep the controller in-process. The accompanying 0.2.0 Helm chart pins the stable app and image versions for cluster installs.

That distinction matters for data and analytics agents. A scheduled metadata audit, warehouse-health check or report-generation job is no longer just a recurring prompt: it becomes a persisted resource with run history, pause and resume controls, and an identifiable owner.

Ownership is part of the runtime

The same release tightens who can see and operate those resources. According to the release notes, users may read sessions, turns, events, metrics, schedules and runs they created, or resources attached to an agent they manage. Mutating a session or schedule remains restricted to its creator. TrueForge also added an optional OIDC email allowlist, agent-level external authorization and mutual TLS for selected internal hops.

The guardrails extend to tools. Request-scoped MCP servers and skills can be supplied without appearing in the shared settings catalog; destructive MCP tools retain approval by default, while read-only selections can clear it. Remote MCP responses are capped at 50 MB to reduce the risk of oversized tool output exhausting the process.

A useful boundary for NL2SQL teams

TrueFoundry’s original TrueForge architecture post presents the harness as the layer around a model: the execution loop, context compaction, sandboxing, approvals and persisted state. Models and MCP servers remain bring-your-own, and the sandbox is invoked as a tool only when code execution is needed.

The company’s own 14-task Enterprise-Bench comparison claimed roughly 11 solved tasks for both TrueForge and Claude Managed Agents, with TrueForge using fewer tokens and costing about 30% less when both used Opus 4.8. That result is vendor-run, uses an LLM judge and is not a text-to-SQL benchmark, so it should be treated as a reproducible claim rather than independent proof.

The stronger signal in 0.2.0 is architectural. For teams operating database agents, the release makes schedules, authorization and observability properties of the harness instead of one-off application code. The remaining evaluation question is whether those controls hold under real multi-tenant workloads and long-running database tasks; the release notes provide the machinery, not that production evidence.

sources

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.

    more from the desk