Day 2 had a theme, and it was AIforce.
The first session was on how the AIforce architecture works, and the part that matters if you are the one signing off is the connection point. AIforce, the layer above the Agentforce agents, connects to whatever integration layer you put in front of it. Claude. Slack. Something you build yourself. That is what putting Headless 360 to work looks like in practice.
That changes the arithmetic for most of the organizations I work with, and not for the reason the keynote would give you. The most common request I get has nothing to do with AI. People want to work with their data without logging in, clicking through to the right screen and punching a value into a field. Text the system an update after a client meeting. Ask it from the road where an opportunity stands. Type the question into Slack and get the answer back in Slack. None of that is exotic, and all of it has been out of reach for any organization without a development team behind it.
The second session was on standing your own org up as an MCP server. MCP is the Model Context Protocol, the standard an outside tool uses to reach into a system on your behalf. What surprised me is that the Claude connector, using Claude itself as the interface, is generally available right now. Not a pilot, not a roadmap slide. Adam and I are testing it against real orgs the week we get back.
That settles something I raised on Day 1. The demo I watched ran out of Claude’s app into Salesforce rather than the reverse, and I said then it needed confirming. Two days in, that is the direction. The interface moves out. The system of record stays where it is.
The structure is close to something we already run in house, a set of Claude agents wired into our own systems, one per business function. The difference is that Salesforce is putting real structure around it, and structure is the part you cannot easily build yourself. It is also the part that decides whether a thing like this survives contact with an audit.
The third session was on securing Apex and simplifying agent access to the data inside it, and the Named Query API was my takeaway. It lets a developer put an agent against real data without the overhead of a pile of Apex to get there, which drops the cost of trying one of these a long way.
Which leaves the same constraint I have been pointing at all year. The interface problem is being solved in public, on stage, by people with more engineers than any of us. Your data and your processes are still yours to fix, and nothing announced this week does that for you.
Adam’s Day 2 dispatch, from the Keyboard, is here.
Responses