Connect:AI San Francisco 2025 was about laying the groundwork for the agentic enterprise. This year, the conversation shifted to what comes next: how to stay in control once that foundation is built.
This shift in thinking set the tone when IT leaders, architects, and product experts gathered in San Francisco for Connect:AI SF 2026 to tackle one of the most prevalent obstacles in enterprise AI today: scaling autonomous agents without losing your grip on them.
Setting the stage
During the opening keynote, Matt DeTroia, EVP of Data Foundations at Salesforce, started by painting a picture of where the industry stands. DeTroia explained that enterprise AI doesn’t operate in a vacuum – it has to work with the systems, data, and integrations companies have spent years building. For most organizations, DeTroia said, the challenge no longer lies in introducing AI. It’s in building the architecture to manage it once it’s everywhere.
3 pillars for scaling agents
Ahyoung An, Chief Operating Officer at MuleSoft, followed up with a more concrete framework for what enterprises need as agent adoption scales:
- Open by design
- Engineer for Trust
- Always in motion
These three pillars shape MuleSoft Agent Fabric, the control plane that gives teams one place to discover, observe, orchestrate, and govern agents and LLMs. From there, product leaders Ashley Hathaway, Karl Albertsen, and Jing Li walked through what each pillar means in practice.
1. Open by Design
This stems from a problem most IT teams know firsthand: agents and automations are being built everywhere, often with no one keeping track of them. MuleSoft’s fix is ​​a set of automated scanners that crawl cloud environments and register their findings into a single catalog. This gives teams a single registry to apply consistent rules across third-party gateways, instead of hunting down rogue deployments after the fact.
2. Engineer for Trust
This is built around two things that keep IT leaders up at night: runaway token spending and data slipping through the back door. Rather than asking companies to rebuild what they already have, this approach layers security on top of it. It’s also smart about costs. Simple requests get routed to lighter, cheaper models, while harder problems receive the support of more capable ones. Behind the scenes, it trims down bulky prompts to cut waste and keeps a constant eye on where your data is flowing, catching risky exposure before it becomes a problem.
3. Always in motion
This is about who gets to build. The goal is to move past requiring specialized coding skills, so developers and business users alike can support agentic workflows using natural language. By putting these capabilities into the tools people already use, MuleSoft is letting users audit risk and adjust security policy without leaving the tools they work in every day.
Live from customers
After the keynote, the focus shifted to a panel of IT leaders moderated by Harsh Karmarkar, AI Strategist at Salesforce. He was joined by Ajay Sabhlok, CIO and CDO at Rubrik; Nikesh Kotecha, Head of Data Science at Stanford Health Care; and Abhijeet Padhan, Global Head of Product and Technology Digital Platforms at HP.
Despite a dynamic conversation, the panel kept circling back to one point: governance has to come first. Skip it, and companies are met with model decay, corrupted data, or a mess of unmanaged permissions, all problems that are harder to fix after the fact than to design around from the start.
Sablok shared how Rubrik broke down internal silos by building a centralized view of the organization’s agent and automation efforts. By establishing this visibility, they were led to an unexpected realization: multiple teams were independently building the same tool to solve the same problem. Once this was made obvious, they consolidated the work, pooled resources, and moved faster as a result.
In addition to governance, safety nets were identified as non-negotiable. Panelists emphasize the need for evaluation frameworks and rollback plans built in from day one, not bolted on after something breaks. As the panel put it, governance isn’t a feature you add later, but rather the starting point for building infrastructure.
The panel wrapped up with the reminder to focus on what actually moves the needle. Rather than chasing novelty or projects that only benefit one team, the advice was to measure agentic initiatives through cost reduction, efficiency gains, and real business outcomes.
Looking ahead
Following the keynote and panel, the day moved into lunch and a set of breakout sessions covering Agent Fabric, headless MuleSoft, and more. Connect:AI San Francisco closed out with networking, giving attendees a chance to continue the conversation with peers.
If last year was about laying the groundwork, this year made clear what’s next: turning that foundation into a system enterprises can actually run on. A system that’s open enough to connect everything, secure enough to trust, and accessible enough for anyone to build on. To see how MuleSoft is helping enterprises put this into practice, explore Agent Fabric and stay tuned for upcoming Connect:AI events.
