Agentforce Labs First Look: Real AI Innovation at TDX 2026

Salesforce just dropped labs.agentforce.com at TDX 2026, and honestly, I haven't been this excited about their ecosystem in a while. I’ve been a massive Salesforce fan for years, but ever since large language models took over the tech industry, I’ve felt a bit underwhelmed by their AI offerings. That changed today. What they just revealed constitutes real, serious innovation.

I decided to dive right in, completely unscripted, to see what this new platform can actually do.

Pushing the Agent Builder

I started by throwing the most ambiguous, unhelpful prompt possible at the new agent builder: "Make me your finest agent. Totally cutting edge and with all the bells and whistles."

Naturally, it asked for more direction. I pivoted to something highly practical but historically complex: an agent that lets users speak in natural language, listens to a phone call, and automatically extracts field data to fill out Opportunities and Contacts.

While the agent builder successfully mapped out a plan, I ran into a classic AI friction point. It generated the first few nodes and just stopped. I had to repeatedly prompt it to "keep going"—a familiar dance for anyone who spends time working with AI agents. But once it finished, it handed me something incredibly fascinating: an Agent Script, alongside a generated username and environment.

Agent Script and Vibe Coding

Agent Script is easily the standout feature for me. It structures the agent's behavior by breaking it down into prompt, reasoning, and deterministic actions. You can explicitly tell an agent to run specific code alongside its reasoning process. Blending deterministic execution with natural language prompting is absolutely the future of development.

I wanted to see how far I could push this, so I took the generated Agent Script and fed it into an external AI coding assistant. I normally prefer Cloud Code, but I was running low on tokens, so I fired up Kimmy. Combining "vibe coding" with actual, serious Salesforce development is a new frontier. Usually, enterprise Salesforce dev is locked down and highly regulated—you can't just let an AI run wild. But these new tools emphasize observability and safety while still bridging that gap.

Labbox, Arc, and Grid

While Kimmy was crunching the code, I explored the rest of the new environment. The platform provided me with a "Labbox"—essentially an Agentforce lab with every single AI feature already enabled out of the box. If you don't have a dedicated org ready to go, this is the perfect clean slate to start experimenting.

I also stumbled across Arc and Grid. Arc automates plan generation for agent development, handling sub-agent orchestration, logic, guardrails, and smoke tests directly in your Salesforce environment. Meanwhile, Grid manages Agentforce workbooks. The whole setup feels like a massive Model Context Protocol (MCP) suite built specifically for Salesforce development.

Spinning Up a Security Agent

Given that I was poking around a brand-new environment (and potentially flashing credentials on screen), I figured it was a good time to test out the Security Agent.

The setup process was impressive. I dug into the security analysis, API analysis, and metric retrieval features, assigning the necessary permission sets to let it run detailed investigations. What really caught my eye here was the model selection dropdown. Salesforce is offering serious firepower, letting you choose from top-tier models like Gemini 3.1, Opus 4.6, and the 5.4 beta.

I left Kimmy hard at work generating the voice-to-CRM integration while the Security Agent ran its audits. The sheer velocity of these tools is staggering, and I’m just scratching the surface. I'll definitely be building more with this soon.

Watch the video

Check out my full unscripted exploration of the new Agentforce Labs below.

Watch the first look at Agentforce Labs - TDX 2026