The data platform is now the agent's office
Block built agents a chat room. Databricks put a governed agent in Slack. Constellation is building a nuclear engineering assistant on the same stack. The agent moved into the data platform.
- Databricks reports Block developed Buzz, an open-source collaboration space where people and AI agents work together, with Unity Gateway providing access to frontier and open-weight models through a single integration, and Buzz is model-agnostic and can be self-hosted. Self-reported m:…1973
- Databricks says the Genie app for Slack is now generally available, and Genie uses the same governed context and permissions in Slack as it does in Databricks. Self-reported m:…1972
- Databricks reports Constellation, which operates the largest nuclear fleet in the United States, is building a governed data platform on Databricks and is developing a nuclear engineering assistant using Databricks Genie. Self-reported m:…1968
- Databricks reports H6.ai uses deterministic assertions and AI judges to evaluate invoice processing outcomes. Self-reported m:…1970
- Snowflake reports Canva replaced static attribution reports with holdout experiments and a unified measurement model built on Snowflake, and its internal audit team transitioned from sampling to continuous, AI-driven assurance with humans in the loop. Self-reported m:…9823
The agent moves into the chat room
Block has built its agents a chat room. Databricks reports the company developed Buzz, an open-source collaboration space where people and AI agents can work together in group discussions, threads and direct messages, with Unity Gateway acting as the model access layer — a single integration for frontier models from providers such as OpenAI, Anthropic and Google, plus open-weight models. Buzz is model-agnostic and can be self-hosted, and Block is experimenting with using a more capable model to coordinate work across a team of smaller, more efficient models. The agent is not a tool you open. It is a colleague in the channel, with a cheaper colleague doing the routine work under a smarter one's direction. m:…1973
Databricks has done the same thing in the other direction, putting a governed agent inside the chat tool everyone already uses. The Genie app for Slack is now generally available, and it uses the same governed context and permissions in Slack as it does in Databricks, so responses respect existing Unity Catalog and source permissions. Users interact with Genie via mentions in channels or direct messages, and can pin a curated Genie Agent to a channel for domain-specific answers. The integration supports actionable workflows, including creating or updating documents and commenting on tickets via MCP connectors, and scheduling recurring analyses — and admins can set responses to be private to the asker as a default for the whole workspace. The agent does not get a new set of permissions because it moved to Slack. It carries the old ones with it. m:…1972
The foundation underneath
The foundation underneath is the story Constellation is telling. The company, which operates the largest nuclear fleet in the United States, is building a governed data platform on Databricks, standardising its data foundation to fix semantic inconsistencies across retail, wholesale and CRM domains. It is prioritising semantic alignment over model selection, using Unity Catalog for lineage and access control and Data Vault modelling for critical, reusable entities — and it is developing a nuclear engineering assistant using Databricks Genie. A nuclear operator is putting an AI assistant on top of its data platform, and the work it is doing first is not choosing a model. It is agreeing on what the data means. m:…1968
H6.ai shows what the governed version of AI automation looks like in the back office. The company automates accounts payable on top of a company's existing ERP — reading invoices, matching them to purchase orders, assigning lines to accounts and checking policies — and it tests its AI pipelines against frozen financial states using Lakebase, where a branch is a copy-on-write copy of a database at a point in time. It evaluates invoice processing outcomes with a dual-judgment system, combining deterministic assertions with AI judges, and plans to use Lakehouse//RT queries for read-only analytics without data duplication. The pattern is the same as Constellation's: the AI is the visible part, and the database branch, the assertion and the judge are what make it safe to run. m:…1970
The pattern
Snowflake's Expedition 2026 event, which gathered customers to talk about scaling AI from pilot to production, tells the same story from the other platform. Canva replaced static attribution reports with holdout experiments and a unified measurement model built on Snowflake. Northern Trust is modernising its finance function by prioritising trusted data and AI. EY is moving tax operations onto a unified tax data hub serving as a single source of truth. Snowflake itself built an assistant named Raven that provides sellers with reliable account insights, and its internal audit team transitioned from sampling to continuous, AI-driven assurance with humans in the loop. The week's message, from both platforms and their customers, is consistent: the agent does not live in the model. It lives in the data platform, under the permissions, the lineage and the semantic layer — and the companies that get the platform right are the ones whose agents get to do real work. m:…9823
What is still unresolved
- Whether model-agnostic agent workspaces become standard or stay platform-specific
- How much semantic alignment is enough before an agent can act
- Whether governed agents in chat tools change how teams share sensitive data
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.