Databricks made the Genie One Model Context Protocol server generally available on September 22 for all Databricks users. The server sits inside Unity Gateway as a managed MCP service and exposes Genie One so any agent can ask questions, pull query results, check progress, and steer responses — grounded in Genie Ontology across tables, documents, and tools.

The pitch is anti-sprawl: Claude, ChatGPT, Cursor, and internal agents share one semantic layer with fine-grained policies and audit logging on every invocation. A companion MCP App can embed Genie’s process with interactive visualizations and ontology citations. Customer use cases called out include slide agents that fetch trusted numbers mid-deck, CS outreach agents that interpret usage drops with ontology-backed telemetry, and coding agents that align logging changes to agreed business definitions.

A September 16 Microsoft Agent Framework deep dive compares two architectures on the same ski-resort demo. Path one: an advisor delegates to specialist agents over A2A, each with its own model loop. Path two: specialists become distributed skills — a description, a SKILL.md procedure, and typed MCP tools — while domain services stay remote and reasoning moves into the parent.

In three paired runs, the skills path used three advisor model calls versus six or seven across A2A components, with lower wall-clock latency in those cache-affected trials (and higher total observed tokens). The author is explicit: MCP does not replace A2A everywhere. Keep autonomous, long-running, or specialized-model work as agents; migrate bounded, read-oriented competences to skills. Skill discovery in the sample still follows a pinned draft of SEP-2640 rather than the unmerged skills/list and skills/get methods.

Palma.ai raised $1.8 million in pre-seed funding led by D11Z, with Plug and Play Ventures, Deel, Scale Now Ventures, and angels from Cisco and Deel. Founded in 2026 by Patrick Eden and Julian Kolbe, the company sits between AI agents and corporate systems to manage permissions for agent-to-tool requests across Claude, Gemini, ChatGPT, and Microsoft Copilot.

The problem statement is MCP-shaped: every tool historically managed its own permissions in isolation. Palma centralizes policies, approved MCP tools, and recurring workflows, enforces rules at runtime, and writes a tamper-evident audit trail. Capital funds engineering and go-to-market as enterprises hand agents real keys.

Silk, a data platform pitched for agentic AI workloads, closed a $45 million growth capital facility on September 23 led by Avenue Capital Group, with Sequoia Capital, S Capital VC, Vintage, Ibex, West Coast Equity Partners, and others. Proceeds expand engineering, neocloud/sovereign/inference-cloud go-to-market, and the enterprise footprint across AWS, Azure, and Google Cloud. Avenue framed the data layer as the bottleneck as agentic AI moves from pilot to production.

Separately, Chicago-based Go.AI (formerly Go Abacus) raised $85 million Series A led by Updata Partners, with GFT Ventures and LAUNCH, bringing total funding to about $90 million. The company says it remains profitable and is pushing The Go1 — on-premises AI hardware plus Go.OS software — into regulated sectors that refuse to move sensitive data outside their perimeter, with a seat-based rather than usage-based pricing model.

Cambridge-based Ekai raised $1.7 million pre-seed led by Misneach with C10 Labs to attack what it calls the “meaning gap” and “context rot” in enterprise AI. The product’s forward-engineering flow starts with domain experts as ground truth for business meaning, then translates into machine-readable semantics with validation against the warehouse.

Ekai positions itself as an augmentation layer, not another catalog: humans drive the semantic model; AI handles mechanical translation. It runs in the customer’s cloud, reads data in place without copying or retaining it, and lists on Snowflake Marketplace plus Databricks, BigQuery, Postgres, ClickHouse, DuckDB, Redshift, and Azure Synapse. Funding accelerates product and go-to-market.

On September 23, OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and Hugging Face’s Clem Delangue told the UN Security Council the industry needs international coordination on capability measurement, risk assessment, safeguards, human oversight, and fast incident reporting. Amodei again framed mismanaged AI as a civilizational risk and said his company will slow releases as needed for safety. The Trump administration’s representative rejected “centralised control and global governance of AI” by international bodies.

In Washington the same day, Sen. Bernie Sanders and Rep. Greg Casar unveiled a bill that would ban artificial superintelligence, pause advanced AI development until federal safety rules exist, create a Department of Artificial Intelligence, and require federal approval before deploying advanced systems — with steep penalties. The bill faces long odds in a Republican-controlled Congress. Separately, reports say Anthropic’s IPO timing has slipped past the November midterms, delaying a public look at frontier-lab economics that underpin hundreds of billions in Big Tech commitments.