Salesforce wants you to believe it has cracked the code on autonomous AI agents. With the release of Agentforce 3, the company promises a leap forward: agents that plan, reason, and act independently, orchestrating outcomes across sales, service, marketing, and beyond. They’ve even introduced a shiny new “Atlas Reasoning Engine” to give these agents brains, not just arms and legs.

But strip away the branding and bravado, and what emerges is a more nuanced, and still constrained, picture: Agentforce 3 is an upgrade, but it remains bound to the Salesforce universe and rooted in deterministic scaffolding. It’s nominally smarter, yes. But is it transformational? Only if you accept Salesforce’s definition of transformation.

What’s New in Agentforce 3

Agentforce 3 ships with some putatively notable improvements:

• Atlas 3.0: A faster reasoning engine that can break down goals, plan execution steps, and choose the right actions using a context-rich memory.
• Command Center: Real-time observability with OpenTelemetry support and dashboards to trace errors, latencies, and agent behavior.
• MCP Integration: Out-of-the-box support for Salesforce’s implementation of the Model Context Protocol, allowing external “plug-and-play” agent interactions.
• Expanded Agent Actions: 200+ prebuilt industry actions plus low-code agent builders, prompt design tools, and analytics.
• Trust Framework: Guardrails, zero data retention policies, and supervised learning loops to (in theory) ensure safe AI deployment.

This is not smoke and mirrors. Agentforce 3 represents actual engineering work. But it’s also an elaborate extension of Salesforce's control plane, not a neutral AI operating system.

Where It Still Falls Short

Autonomy Still Has a Ceiling

Yes, Agentforce 3 agents can reason, plan, and execute. But their “freedom” is still highly constrained:

• You define the topics.
• You define the actions.
• Atlas just selects the best combination and executes them.

There’s no emergent behavior. No agent invents new skills. No reinforcement learning. This isn’t “autonomous AI”, it’s selective automation with decision support.

Verdict: Smarter flows, not true agents.

Governance Is Better, but Still a Black Box

The new Command Center improves observability, with live traces, telemetry, and dashboards. But explainability — why an agent made a decision — is still lacking. You can watch it happen, but not fully understand it. And supervised feedback loops still rely on manual tuning.

Verdict: You can trace the what, but not always the why. Note: this has compliance implications for those in regulated industries.

MCP is Progress, Not a Panacea

The addition of Model Context Protocol (MCP) support is Salesforce’s answer to the interoperability challenge. It theoretically enables agent-to-agent interactions across platforms (e.g., AWS, Google, Stripe). But in practice, most customers still report success when staying inside the Salesforce walled garden. Custom MCP endpoints? Still early. Real portability? Not yet.

Verdict: Openness in theory, stickiness in practice.

Vendor Lock-In Is Alive and Well

Agents live in YAML templates, stored actions, Flow configs, Apex snippets, and Salesforce’s metadata model. Even with MCP, if you walk away from Salesforce, you walk away from the brain and the body of your agent ecosystem.

Verdict: You’ll love Agentforce 3 most if you’re already all-in, but who is really all-in?

Still Fragile at Scale

Real-world users report limitations:

• Max 20 active agents per org
• ≤10 recommended actions per topic
• Latency and hallucination spikes with messy data
• Unpredictable behavior in non-English contexts

Agentforce 3 works beautifully in Salesforce’s curated demo environments. In messy, multilingual, multichannel real-world contexts? It still needs diapers.

Verdict: Great in the lab. Mixed in the wild.

Bottom Line

Agentforce 3 is an advancement of Salesforce automation. It introduces a real reasoning engine, better observability, improved trust tooling, and a shot at interoperability. If you are a current Salesforce customer it deserves attention.

But let’s not pretend it’s a revolution. Behind the scenes, it’s still Salesforce-flavored determinism, just with better wrappers and some LLM-based sugar. It’s a governed, AI-powered automation suite, not a platform for general-purpose or even marketing-focused intelligent agents.  Remember: the most high-value agentic use cases entail cross-platform orchestration (this recorded webinar, Get the Real Story on AI Agents for MarTech, explains more).

If you’re already deep in the Salesforce ecosystem, Agentforce 3 begins to give you a cleaner, faster way to scale intelligence across your stack. If you're looking for open agentic infrastructure?

Keep looking.

Ready to get your Agentic Strategy moving in the right direction? We have a lot to offer.