Hey there 👋 I read the Salesforce announcement twice, because the first time through I thought it was just keynote swagger. The pitch, in plain words: you stop coming to Salesforce, and Salesforce comes to you, inside Claude or Slack or wherever you already work. Salesforce sells screens. Tens of thousands of them, the tabs and buttons and fields a whole trade of admins gets paid to arrange. And there it was, telling you the screen is the part it now wants to make optional. Then I lined it up against the rest of the week and the swagger started to look like a plan. UiPath rebuilt its integration layer so a coding agent can call it directly. Riverbed put an agent in front of the network dashboard so you ask instead of dig. ByteDance went the other way in China and shipped a version of Feishu with agents as first-class citizens, keeping them inside its own app. Four companies, one buried question: where does the agent sit, and whose front door does it walk through. My bias, for what it is worth: the front door is the real fight now, and a lot of companies have not clocked that they are in it. If your product is something a person opens and clicks through, you are about to learn whether it survives as something an agent calls instead. That is a harder question than which model tops the benchmark, and this was the week it landed on the table.
The Big ThingSalesforce moved its whole platform to the agentAt Dreamforce this week Salesforce unveiled AIforce, a live interface layer that takes the data, workflows, business logic, permissions, security, and governance locked inside Salesforce and exposes all of it to an outside AI interface. It launches with three front ends the company calls Claudeforce, Slackforce, and Agentforce Coworker. Salesforce's own framing is blunt: people do not come to Salesforce to get work done anymore, Salesforce comes to them, in Claude, Slack, Lightning, or wherever they work. The engine underneath is the Headless Toolkit, an open architecture that publishes Salesforce as MCP servers, APIs, plug-ins, skills, and developer tools. So an agent sitting in Claude or Slack can read a record, run a workflow, and take an action while the platform's permission model and audit trail ride along with every call. You do not rebuild the rules inside each agent. They travel with the request. For a category that spent fifteen years training people to come to its screens, that is a real turn. This is not a roadmap slide either. Agentforce Coworker turns on with one button, no migration and no new permission model, and Salesforce says 100,000 users switched it on in the first 35 days. The Adecco Group announced a rollout across more than 40 countries after a pilot in the UK and France, running on Anthropic's Claude across sales, recruiting, and candidate work. Whatever you make of the framing, the deployments are real and already moving. Read the numbers for what they are. The 100,000-in-35-days figure and the customer logos are Salesforce's, straight from the launch, so grade them like marketing. The Adecco rollout is a separate, company-issued announcement, which makes it a firmer signal than a keynote stat. And "Salesforce comes to you" is positioning, not a product you can buy by that name. What is concrete today is the Headless Toolkit and the MCP and API surface you can start building against. Ship it? If you run Salesforce, treat this as a build task rather than a keynote to admire. Stand up one read-and-act flow through the Headless Toolkit, say updating an opportunity stage from Slack, confirm the permission model and the audit trail hold under an agent, then widen from there. If you do not run Salesforce, the pattern is still the memo: the biggest CRM on the planet just bet that the winning interface is somebody else's agent, and your software is on the same clock. Sources: Salesforce Newsroom "AIforce" announcement and SiliconANGLE, both Sep 15 2026; Adecco Group press release, Sep 15 2026. The 100,000-user and customer figures are vendor-supplied and graded as such.
Tour de Headlines🔌 UiPath made itself the governed way in for coding agents. A coding agent can write an integration in seconds and still have nowhere safe to run it. UiPath's answer is its Integration Service, which it now positions as the agent-ready connectivity layer underneath. The agent builds the connector; UiPath holds the credentials, the triggers, and the scoped access to your ERP, CRM, ITSM, and legacy systems, so the agent reaches real business systems without a human hand-wiring each brittle link. It builds on the coding-agent integration UiPath shipped in May, which started with Claude Code and OpenAI Codex. Same move Salesforce is making, from the automation side: quit asking the agent to learn every system, and hand it one governed door. If you run UiPath, test the coded path on a staging system and confirm secrets stay centrally managed before it touches production. 🌏 ByteDance took the opposite bet, and it is worth watching. On September 16 it shipped Feishu 8.0 with AI agents billed as first-class citizens, wired into its Doubao Workspace, and it folded the whole Feishu team into its AI division. Where Salesforce and UiPath open themselves up so an outside agent can call in, ByteDance is keeping the agent inside its own app and betting you will come to it. That is the same question from the other end: does the agent travel to your software, or does your software become the place the agent lives. Fair warning that this is a read on strategy rather than a tool you deploy tomorrow, and its sharpest edge is China's office-suite fight with Tencent and Alibaba. Still the cleanest illustration this week of the two roads on offer. 🩺 Riverbed handed the network dashboard to an agent. More on this below in Tool of the Day, but the short version: Network 360 puts an agent in front of your packet and flow data so a NetOps engineer asks a question instead of clicking through six screens to build the answer. It shipped September 16 and lands inside the AppResponse and NetProfiler tools those teams already run. The interface shift the whole issue is about, showing up in an unglamorous corner of ops.
Sponsor Your team's calls are full of signal. Measure it. RapportScore reads the calls your team already records and scores how your people communicate, then turns it into coaching they can use. Stop guessing whether a rep is building trust or talking past the buyer, and see it in the numbers. See your team’s score → |
Tool of the Day🩺 Riverbed Network 360Ask your network what broke instead of digging through dashboards for it. If your team runs NetOps, this one is deployable today and it slots into tools you already run. On September 16 Riverbed launched Network 360, folding its agentic AI, Riverbed IQ and a conversational layer called Q, into its AppResponse and NetProfiler products, with NPM+ extending the view across remote users, Zero Trust, and public cloud. The agent correlates packet and flow data, names a likely cause, and recommends a next step, and an engineer can open the underlying evidence to check the call before acting on it. IQ and Q land as native capabilities inside AppResponse and NetProfiler, the tools NetOps teams already run, so it is not a separate product to bolt on. Deploy it narrow: wire it into the runbook for one service boundary, measure whether its recommendations cut escalations, and keep an eye on false positives before you widen it. Read the Riverbed launch
Worth a Click- Salesforce AIforce. The Dreamforce announcement in Salesforce's own words: the Headless Toolkit, the Claudeforce and Slackforce front ends, and the case for making the interface an agent. (salesforce.com)
- UiPath Integration Service. The agent-ready connectivity layer itself, and how activities, triggers, and governed connections let a coding agent reach enterprise systems without hand-built glue. (uipath.com)
- Linux Foundation welcomes TRACE. A new open spec for hardware-attested, verifiable runtime evidence for AI workloads, so you can prove what an agent ran and what data it touched. The trust layer forming under all of the above. (linuxfoundation.org)
Call it the surface war. Strip the logos off the week and the fight is not about which model is smartest, it is about where the agent sits relative to your software. Salesforce opened its whole platform so an agent in Claude or Slack can drive it, and said out loud that the screen is now optional. UiPath made itself the governed door an agent walks through to reach your systems. Riverbed slid an agent in front of the network so you ask instead of dig. ByteDance bet the other way and pulled the agent inside its own app. Every one of them is answering the same question a different way: when your customer stops opening your product and starts asking an agent to use it for them, are you the thing it calls, or the thing it skips. That question was abstract a year ago. This week four companies started spending real money on their answer. See you Monday. — Ron |