Agentic AI

TDX 2026 Roundup: Agentforce Edition

Open source, faster access, and a full lifecycle for building agents with Agentforce.

Susannah Plaisted

April 15, 2026 5 min read

Our forward-looking statement applies to this blog.

At TDX 2026, we’re focusing on a question we keep hearing from developers: why does it take so long to go from an agent idea to something running in production? The gap between a working prototype and a deployed agent is real — provisioning, authoring, testing, deployment, observability. Each step has its own friction.

This year we shipped concrete answers. We open sourced Agent Script, launched Agentforce Labs for rapid setup along with other product previews, and released ADLC skills that close the loop from sandbox to production. Here’s what each one does and why it matters.

Open Sourcing Agent Script: An Agent Definition Language for the Enterprise

Agent Script is an agent definition language that lets developers specify when agents should use LLM reasoning and when they should follow deterministic logic. Subagents, actions, variables, guardrails, transitions—all defined in structured, strongly-typed files that coding agents can work with natively.

That’s the key point. You’re not going to write Agent Script by hand. Agents will write it for you. Claude Code, Cursor, Codex—coding agents are writing the majority of code now. If they have Agent Script skills, they get better at building agents. If they don’t, developers write everything manually.

But all this new innovation extends beyond writing net new code. Indeed is an early proof point. Their team uses coding agents to scale agent testing, proving the same skills developers use to build agents also work to validate and improve them.

In opening up the authoring layer, we’re giving developers direct control over tool calling, sub-agent routing, and variable storage. As models get smarter, we intend to take those learnings and further invest in composability across the platform.

A standard also needs room for specialization. Agent Script works as a base language with platform-specific extensions—what we call dialects. Think of it like SQL: the core language is standard, but Oracle, Postgres, and MySQL each add extensions for their specific capabilities. The Agentforce dialect defines how agents work. Mulesoft Agent Fabric extends the language to orchestrate across hybrid integration landscapes. The base language remains open; the extensions reflect the platforms they run on.

At TDX 2026, we open sourced Agent Script authoring. The full language specification, grammar, parser, and compiler are available at github.com/salesforce/agentscript. Our goal is a fully open ecosystem for the language, but we’re getting there in phases. Today, we welcome bug fixes, developer tooling improvements, and suggestions on language direction. Over time, we may consider opening more of the ecosystem to enable a more broadly governed agent harness community. Dialect capabilities and the base language specification are governed by Salesforce — that’s how we ensure enterprise reliability as the language evolves.

Agentforce Labs: Innovation at the Speed of AI

Open sourcing Agent Script removes friction from authoring. But friction also shows up before developers write a single line of code — in provisioning, setup, and configuration. If we want developers building on the platform, we have to meet them where they are.

Agentforce Labs is an incubation program for tackling these friction points. We ship experiments, test them with real developers, and either graduate them to core Agentforce or retire them. We don’t wait for GA release cycles to see if an idea works.

The first experiment is Agentforce Labs Quickstart: instant access to Agentforce from the IDE of your choice. No org provisioning, no setup screens. Developers who want to kick the tires of Agentforce can connect and start building in Claude, Codex or Agentforce Code. More experiments will follow at labs.agentforce.com as we identify and eliminate the next friction points.

The Full Lifecycle: From POC to Production

Agent Script opens authoring. Labs opens access. The last layer is everything between building an agent and running it in production — testing, safety reviews, deployment, and observability. That’s where ADLC (Agent Development Life Cycle) skills come in.

ADLC skills close the loop from sandbox to production and back. Developers author and validate `.agent` files locally, run LLM-powered safety reviews that catch harmful patterns keyword filters miss, deploy without leaving their IDE, and feed production session traces back into their authoring workflow. The cycle is continuous: build, test, deploy, observe, improve.

The skills work in any environment that supports the skill.md standard. The repository is available at github.com/SalesforceAIResearch/agentforce-adlc with installation instructions and full documentation. It’s also available in Agentforce Code and Agentforce Studio.

What Else We Announced at TDX 2026

Agent Script, Agentforce Labs, and ADLC skills are part of a broader set of announcements under the Agentforce Developer Platform:

Agentforce Vibes 2.0 – An AI development environment that understands enterprise metadata through Salesforce Unified Catalog, supports multiple models (Claude Sonnet, GPT-5), and generates code across frameworks like React, LWC, and Apex.
Agentforce Experience Layer (AXL) – Define interactive components once and embed them into conversational flows across surfaces (web, mobile, Slack, voice) without rebuilding for each platform.
Expanded AgentExchange – Trusted marketplace to discover, deploy, and monetize apps and agents, now with 10,000 Salesforce apps, 2,600+ Slack apps, and 1,000+ agents, tools, and MCP servers from across the ecosystem.
Agent Script for Voice – Agent Script now works across text and voice channels. Same deterministic control over voice interactions as text — build once, deploy everywhere.
Agentforce Mobile SDK – Deploy Agentforce natively in iOS and Android apps with push-to-talk voice, so customers can speak to get help without leaving the app.
Session Trace OTel API (beta) – A single REST call returns a complete Agentforce session trace — interactions, audit events, and feedback data — as standard OpenTelemetry (OTel) spans. Export directly to Splunk, Datadog, New Relic, or any OTel-compatible collector with zero transformation.
A/B Testing API (pilot) – Run controlled experiments on Agentforce agents via API. Split traffic between agent versions, measure results, and promote the winner.
Testing Center Enhancements (GA – May) – The next generation of Testing Center in Agentforce Studio adds custom scoring evals (define what “good” looks like for your use case — brand voice, compliance, resolution quality), multi-turn conversation testing with voice simulation, and a flexible UI with run history.

Companies like Adobe, Asymbl, Engine, Indeed, Paypal, and Petaluma Creamery showed how they’re deploying agents at scale in case studies and keynote demos.

Start Building

The authoring tools are open. The experiments are live at labs.agentforce.com. If you’ve been waiting for the right moment to build on Agentforce, this is it.