From Autonomy to Accountability: How to Think About Trust in the Multi-Agent Future
AI agents can identify unsafe actions in reasoning but still execute them anyway
“We need a hybrid approach, combining the LLM's probabilistic reasoning with hard-coded, deterministic guardrails that exist outside the agent's internal reasoning loop.”— Salesforce
Salesforce convened 21 organizations across industry and academia to produce a white paper on trust in multi-agent systems, surfacing a critical finding: agents can recognize a request as insecure in their internal reasoning yet still execute the harmful action — a 'thought-and-action disconnect' that cascades across connected platforms. The paper argues Zero Trust is the baseline requirement for agentic architectures and calls for deterministic guardrails that live outside the agent's reasoning loop. This reframes enterprise AI security from perimeter defense to continuous, per-interaction verification.