Why the Harness Is Bigger Than the Coding Tools Make It Seem: Salesforce Platform Chief Rohan Kumar Unpacks Headless 360

When Salesforce launched Headless 360 at TDX in April, the tooling pointed at people writing code. Sixty or so Model Context Protocol (MCP) tools and thirty pre-configured coding skills planted the foundations for connecting coding harnesses to Salesforce backends. This week, Salesforce took the next step by surfacing these capabilities for business users with a harness aimed at people who run marketing, sales, field service, and storefronts. Top-line capabilities include the new or enhanced Headless 360 MCP server, Data 360 MCP server, Agentforce Coworker, Slackbot integrations, and roughly 100 pre-built skills. Rohan Kumar, President and Chief Platform and Engineering Officer at Salesforce, sets out the vision: > The vision really over here is: how do we make it really easy for agents to augment practitioners in their day-to-day work? So far, if you take a look at it, agents could observe what's going on, but the vision that we have with a headless approach is how do you take the entire value that we would normally create with user workflows, data, identity, security, and make that available as programmable APIs that agents can read and actually get the work done, which I think significantly unlocks the value for enterprise practitioners. > > > So that's the uber vision, and it's not just about an MCP endpoint and a set of tools associated with it. That's a part of it, but it's really how do we think about the entire Salesforce application and platform ecosystem that's made available through a set of APIs that business users and developers can use as they build out agents that actually help them with their day-to-day work." Headless 360 has spent four months as a developer story. This expansion widens it to the business users who will approve and run these agents. Kumar frames the shift as a harness around the model, extending the same idea that has been growing up around coding tools. That harness helps subject matter experts and front-line users connect emerging LLMs with Headless 360 infrastructure that manages enterprise context, decision-making, execution, data governance and identity. One key consideration has been extending the trust foundation enterprises have already built across their Salesforce estate to business users who might not be familiar with the security and governance implications of approving new agentic workflows. Fully headless Kumar joined Salesforce two months ago from Microsoft, where he spent 28 years, most recently as corporate vice president of Microsoft Security and before that in Azure Data and SQL Server. He took on the engineering side of his current title earlier this month. The foundational piece is the Headless 360 MCP server, now in open beta, which Kumar describes as taking the entire Salesforce application and platform ecosystem and making it available through a set of APIs callable from Agentforce, Claude Cowork, Claude Code, ChatGPT, Microsoft Teams and Slack: > The vision really over here is to meet the practitioners wherever they work. So we're not being very opinionated about you have to come into this place or a particular interface to actually get your work done. It's a very open and very composable ecosystem. A supporting piece is the Data 360 MCP server, with about 200 APIs today. This supports a data and knowledge layer designed for contextualization, efficiency, governance, and cost management. Kumar explains: > Through a technology called zero copy, we are able to reach into multiple data stores, warehouses and lakehouses across your enterprise and synthesize a very common knowledge layer, which essentially makes the agents a lot more efficient and a lot cheaper to run. So if you think about the tokens that the agents burn, because of the investments that we've done in Data 360, which is now available through this headless interface for agents, that actually makes it extremely rich and promising. Also shipping are agentic user experiences surfaced through new integrations and front ends. The headline is Agentforce Coworker, which Kumar frames as both a product for Salesforce admins and a demonstration piece for how agents can be built on top of the headless interface. Others include a new Slackbot, the first services from MuleSoft and Informatica, and broader support for third-party agentic harnesses. Kumar says: > For the most part, I think we are actually fully headless, but as you can imagine, there's innovation that's happening every day in our core applications and in some of the work we've done with Tableau, MuleSoft, Informatica, Data 360. But what I can tell you is that headless is the key pivot. As in, with any new innovation we get, we always think: how do we get this from a headless standpoint... It's not just about getting access to the data, but how do we code up the workflows to be agent-readable through APIs? According to Salesforce, customers are already building on it. Engine, a travel platform for small and medium businesses fielding more than 530,000 support requests, launched its EVA support agent in 12 days and now resolves half of all customer chat interactions without a human on the line. Job site Indeed dropped its browser UI dependencies and cut agent build times from months to weeks. BACA Systems, a 50-person manufacturer whose single administrator manages more than 1,300 custom objects and 700 flows, consolidated inventory, procurement, receivables and field service onto Headless 360. VentureCrowd, an Australian investment platform working under financial services regulation, took its deployment cycles from two months to two weeks. Extending the harness Harness has mostly been a word in the coding world. It describes the scaffolding wrapped around an LLM that turns it into something useful, including the build environment, runtime, loops, and quality control. Kumar wants it to carry considerably more weight: > Typically, when people talk about harness, it is a word that I think is sort of evolving on its own. Initially, it was basically the building and the runtime for the agents, which included things like observability and monitoring. Our vision, and we'll be sharing a lot more about this, is much more expansive than that. For us, there is this notion of context, which is essentially all your enterprise understanding that's there. Then there is typically what people look at as the builder harness, which is what we call the agency, where you're effectively using the model, using that enterprise context to basically make decisions that enable you to get to an outcome. > > > Then there is how do you execute the action, which is once you've made the decision, how do you execute the action, and we have workflows for that through the Salesforce apps, through Data 360. And then finally it's around the data governance piece, which is as you're building out the knowledge layer, how do you know what changes were made across what data stores? Who made those changes? Can you trust those or not? And then finally, the security piece, which is around things like agent identity, data security. One practical payoff is extending the deterministic workflows enterprises have spent years building in Lightning, the Salesforce UI, into new conversational and autonomous ones. The headless APIs capture constraints an admin has already set through skills. As Kumar puts it, the data, the knowledge and the enterprise workflows and constraints together are what agents need to execute; otherwise, they have to recreate value that has already been built. Agentic governance Salesforce has been working to simplify the governance side of these workflows for business users who might not have the mental models to understand the security and compliance implications of approving new agents. Kumar says: > [Agents] basically run on behalf of a user, and the industry is getting to a point where, as these agents become more and more sophisticated, they start accessing more and more important and confidential d