Agents Must Earn Autonomy: Lessons from Salesforce Agentforce and MuleSoft

Agents Must Earn Autonomy: Lessons from Salesforce Agentforce and MuleSoft

Episode summary At Dreamforce 2026, Salesforce put deterministic control at the center of its agent strategy. Rather than trusting AI models to make every call, Agentforce now pairs LLM reasoning with scripted workflows and hard rules, while MuleSoft’s governance layer enforces identity, tool permissions, and rogue agent detection at the gateway.

Brian, Campbell, and Michael explore what this shift means for enterprises deploying AI agents. Their shared conclusion: autonomy shouldn’t be the starting point. Agents begin inside defined limits and earn more freedom as they demonstrate they can be trusted.

In this episode:

• The deterministic turn. Why one of the largest enterprise software vendors is wrapping probabilistic AI in rules that execute the same way every time, and what that admits about agent reliability today.

• MuleSoft as the control layer. How identity propagation, tool allow-lists, and anomaly detection keep agents inside their intended scope, even if the model is manipulated or confused.

• Consistent isn’t the same as correct. Deterministic rules remove model unpredictability, but a badly written rule fails reliably and at scale. Governance has to cover the rules, not just the agents.

• Ownership and accountability. Who writes the rules, who reviews them, and who answers for them when something goes wrong.

• Concentration risk. What it means when a single vendor owns the data platform, the governance middleware, and the agent runtime.

• Earning autonomy. A practical way for leaders to expand agent independence step by step, based on evidence rather than vendor promises.
  

Key takeaways

1. Agent reliability comes from the controls around the model, not the model alone.

2. Deterministic governance shifts risk from model behavior to rule design. Someone has to own that design.

3. Keep governance independent from the systems it governs wherever possible.

4. Grant autonomy in stages, tied to demonstrated performance and clear accountability.
   

HostsBrian Silverman, Campbell Robertson, and Michael Muhlfelder