In this briefing
- 01Office software moves Agents from answers to editable work objects
- 02Plugin directories start to carry release, review and adoption feedback
- 03Spreadsheets bring Ontology, permissions and audit into the operating surface
- 04Agents in team chat need first-class identities and reveal an early market boundary
- →What to watch next
- ↗Sources and verification
- Microsoft puts editable Office documents, small applications that can be generated and continuously running Agents into one Copilot experience, and describes Autopilot as having its own identity, memory, computer and workspace. The release states of Home, Code, Managed Runtime, Autopilot and related Teams capabilities differ; most are still rolling out in stages or remain in preview.
- Anthropic's new portal accepts a remote MCP connector or a package combining MCP and Skills, which is published after automatic validation, a safety scan and review, and then receives installation and search analytics. Directory review can form a distribution gate, but cannot substitute for an enterprise's independent validation of a third-party plugin's permissions, data flows and runtime behaviour.
- Databricks plans to connect Row Zero's governed spreadsheet to Genie, while Ando brings Agent identity and context into team messaging. The former has not fully disclosed its integration schedule or performance boundaries; the latter still mainly serves small teams, so neither supports an inference of broad production adoption.
Office software moves Agents from answers to editable work objects
On 25 September, Microsoft announced Home, Code and Autopilot for the new Copilot. Home puts Chat, Cowork and editable files in Word, Excel and PowerPoint within one experience. Code lets people create an application, tracker, dashboard, automation or workflow in natural language; Microsoft says that it runs in a sandboxed environment and can be hosted inside an enterprise tenant. Copilot Managed Runtime is described as IT-governed hosting infrastructure for applications produced through Cowork, Code and Copilot Studio.
Microsoft defines Autopilot as a continuously running cloud Agent with its own identity, memory, computer and workspace, able to appear in Teams, Outlook, channels and documents. This does not mean that Home and Code are broadly available: both will first roll out through the Frontier programme. Managed Runtime is currently in preview. Autopilot is scheduled to expand its private preview at the end of the month. The related Teams capability also remains in private preview. Fabric IQ is GA for Chat and Cowork, but its Code integration is still due through Frontier.
For enterprise adoption, the object under control is expanding from one conversation to persistently editable files, executable applications and work spanning several sessions. Identity, tenant boundaries, sandboxes, version records and human takeover therefore need to be tied to the particular work object. Microsoft's announcement is a product roadmap with staged availability; it is not evidence of cross-sector production reliability or returns. Preview capabilities, GA capabilities and future integrations remain separate states.
Plugin directories start to carry release, review and adoption feedback
On the same day, Anthropic opened the submission portal for the Claude directory. A developer on a paid Claude plan can submit one remote MCP connector or a plugin package hosted on GitHub and composed of an MCP server and Agent Skills. A Claude Code plugin can also contain LSP components, commands, hooks and agents. A submission is automatically validated and safety-scanned; the developer can inspect review progress and feedback, then choose when to publish after approval.
After a plugin goes live, the portal provides installation data by product surface and version, together with listing views and search-term analytics. Anthropic also states that Claude supports the latest specification commonly called MCP 2.0 and its stateless core, and refers to MCP Apps and Enterprise Managed Auth. The unified discovery experience will still roll out over the coming weeks. Scanning and review show that the directory has a publication process; they cannot prove that every third-party plugin meets a particular enterprise's data, identity and compliance requirements.
For enterprise adoption, connectors and Skills are no longer only integration components in a code repository; they are also acquiring submission, review, release, version and adoption-feedback paths. A platform directory can reduce discovery and distribution costs, but permission scope, the OAuth principal, data egress, tool side effects and deactivation remain within the deploying organisation's control boundary. Directory approval is not the same evidence as admission to a business system.
Spreadsheets bring Ontology, permissions and audit into the operating surface
On 24 September, Databricks announced its acquisition of Row Zero and plans to integrate the governed spreadsheet interface natively with Genie. Its material names Genie Ontology, Unity Catalog and Unity Gateway as the shared foundation for business context and governance: queries follow the user's permissions, data refreshes from authoritative live sources, exports can be restricted, results can be written back and interactions can be audited. Finance, operations, sales and marketing teams can still work with familiar formulae, pivots and spreadsheet syntax alongside an Agent.
The acquisition shows a platform bringing an enterprise semantic layer and governance into the data surface used in everyday business, instead of first exporting data into a copy detached from source controls. The status boundary is equally explicit: the official announcement uses future wording including 'will add' and 'will integrate', and does not give a complete integration schedule. The claim of interactive speed across billions of rows remains a vendor statement, with no independent benchmark in the current material. TechCrunch separately confirmed the acquisition and reported that the deal terms were not disclosed.
For enterprise adoption, spreadsheet cells, formulae and model results can become interpretable Agent action records only when entity definitions, metric definitions, permissions and data lineage are consistent. An Ontology and catalogue can provide shared semantics, but they do not automatically eliminate conflicting business rules, historical copies or write-back risk. Completion of the acquisition is not evidence that the product integration has already been delivered.
Agents in team chat need first-class identities and reveal an early market boundary
Also on 24 September, Ando announced a team messaging platform for people and Agents and US$20 million of pre-seed and seed financing. The company says that an Agent can join channels, threads, direct messages, groups and live conversations with an identity, permissions and shared context, while retaining persistent context. Rather than treating the Agent as a one-off bot, this design places it inside daily team co-ordination and makes who can see what, when it may speak proactively and how accountability is recorded into foundation-level product questions.
The public evidence remains at an early stage. Ando's descriptions of product behaviour and outcomes come mainly from the company itself. TechCrunch independently confirmed the launch and financing, and reported customers across several industries in 15 countries, but said that many of the teams are currently small. The financing amount cannot substitute for revenue, retention, production stability or large-enterprise adoption data. It also does not prove that a separate new platform will replace existing collaboration suites.
For enterprise adoption, an Agent in team chat can encounter unstructured decision context, working relationships and tasks spanning several systems. Its identity lifecycle, channel membership, message retention, threshold for proactive behaviour and human-escalation route all need explicit records. An independent collaboration platform demonstrates one product form, but the distance between early small-team use and governance in a large organisation has not yet been validated.
What to watch next
- Whether Microsoft will publish formal technical documentation for tenant isolation, the Agent identity lifecycle, sandbox boundaries and cross-application audit as the Copilot capabilities move from Frontier or preview to GA.
- Whether the Anthropic and Microsoft plugin directories will provide comparable permission declarations, version locking, withdrawal mechanisms and enterprise approval records, rather than only discovery and installation data.
- Whether the Databricks Row Zero integration and Ando team messaging platform will disclose write-back controls, audit completeness, stability, adoption scale and customer-verifiable cases from large production environments.
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing