Golden Data

Issue #11 · July 19, 2026

The product stopped being a place.

Paul BrownWrites Golden Data from Ogden, Utah5 min read

Every Sunday I look for the release note a vendor writes when it thinks it is describing a feature and is actually retiring an assumption. Salesforce published one on July 14.

The Headless 360 MCP Server entered beta in early July with roughly 100 skills behind it — user management, Apex trigger deployment, event-driven integrations, named credential creation. The skills are not the interesting part. The interface is. Rather than exposing thousands of tools, the server presents four: Discover, Describe, Dispatch, and Dispatch Read Only.

Salesforce spent two decades building the most-clicked screen in enterprise software. The version it built for agents has four verbs.

Now look at what those four verbs kept. Every transaction runs as an authenticated user with CRUD, field-level security, sharing rules, and permissions enforced exactly as they are on the screen, plus optional tool-level approval on top. The buttons turned out to be disposable. The permission model did not.

Three other vendors made the same trade the same day.

What’s actually shipping this week

Salesforce shipped an interface with four buttons. Headless 360 went to beta in early July and was written up on July 14, building on the hosted MCP servers that went generally available in April 2026. Discover runs semantic search across APIs and skills, Describe returns specs and dependencies, Dispatch executes with access controls, Dispatch Read Only executes the safe half. A hundred skills at launch. The company whose product has always been the screen shipped the version with no screen, and the one thing it carried across intact was who is allowed to do what.

Clari + Salesloft turned the call archive into a feed. Salesloft Conversation Intelligence became generally available on July 14. Call themes, objections, sentiment, and next steps stop living in a review destination and get embedded into prospecting, pipeline management, and deal execution as a live signal layer — one that feeds agents, not just managers. It builds on the Spring 2026 integration’s MCP server, which opened live revenue intelligence to any AI tool that asks. Chief product officer Kylie Fuentes drew the line herself: the advance is not AI applied to call data, it is conversation data sitting inside engagement, forecast inspection, and outcomes. For a decade the buyer of conversation intelligence was a sales manager who wanted to hear a call. The listener is now a process.

Glean made seven of those archives into rows. On the same day, Glean expanded its native sales connectors to pull context from Clari + Salesloft, Crayon, Fathom, Granola, Otter, Showpad, and Zoom, under one governed assistant. Read that list slowly. An enterprise revenue platform and a free note-taker sit in adjacent rows of the same table, because at the index layer they are the same kind of object: a source of call text. When your product becomes a connector, your competitive set stops being the vendors you meet in bake-offs. It becomes whoever else got a row.

SnapLogic sold the governance, not the speed. SnapCode and the SnapLogic MCP Server hit general availability on July 7 and were announced July 14. SnapCode generates production-ready integrations from natural language inside Claude Code, including the VS Code extension; the MCP server gives agents governed programmatic access to deploy, execute, and manage those integrations against more than 1,000 pre-built connectors for SAP, Oracle, Snowflake, Workday, and ServiceNow. CTO Jeremiah Stone’s framing is that builders should not have to trade enterprise governance for AI coding speed. Note which half is the product. Speed is what the coding agent already had for free.

Four vendors, one Tuesday, four announcements that read as though they are about AI. None of them is. Each took a product whose value was assumed to live in its interface, removed the interface, and found the asset underneath was the rule about who may read what. Salesforce kept sharing rules. Glean sells a governed index. SnapLogic sells permissioned execution. They all sold the right to say no.

What I’d ship in your app this week

Your product is going to be called by something that is not a person, sooner than your roadmap says. Two builds, two weeks each.

One: the four-verb capability surface. Do not expose your REST API to agents and call it done. Copy the Headless 360 shape — a small fixed toolset over a searchable capability catalogue.

  • User moment. A customer’s agent needs to do something in your product at 2am and there is nobody to click.
  • Data shape. One capability row: (capability_id, name, description_embedding, input_schema, side_effects, required_permission, read_only), generated from the handlers you already have.
  • System shape. Four endpoints — discover, describe, dispatch, dispatch read-only. Every dispatch runs your existing authorization path as the calling user. No new permission code. If you are writing new permission code, you have built a bypass.
  • Latency and cost. Discover is a vector lookup under 100ms. Dispatch costs what your handler already costs, plus nothing.
  • Eval shape. Sixty tasks phrased the way a customer would phrase them, scored twice: did the agent find the right capability, and did it get refused when it should have been. The refusal half matters more.
  • Instrumentation. Log discover-to-dispatch conversion per capability. Capabilities discovered and never dispatched are described wrong, not missing.
  • Two weeks in. An agent completes a real multi-step task and a permission check refuses one of its steps. The credible failure is a catalogue that mirrors your database tables instead of your customer’s verbs.

Two: the non-human read ledger. Tag every request with caller type — human session, first-party agent, third-party agent, connector sync — and report the mix weekly. One column and a dashboard. The day your usage curve decouples from your seat count is the day your pricing stops describing your business, and Glean’s connector list is a preview of who will be reading you. The credible failure is discovering most of your reads are already machines and your rate limits were written for humans.

The vendors who moved on July 14 all concluded the screen was not the asset. Go find out what your product is when nobody is looking at it.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown