It's like a spell-checker for AI decisions—catching dangerous or nonsensical agent actions before they happen and helping the agent fix itself.

Salus's core engineering innovation is a hybrid guardrail engine that intercepts every agent action at runtime via an API wrapper. Policies are authored in YAML, Markdown, or plain English and compiled into runtime checks that blend deterministic rule-based validation (e.g., "never issue a refund over $500 without manager approval") with LLM-based semantic evaluation (e.g., "does this email response accurately reflect the customer's complaint?"). Each agent run maintains an evidence cache—a store of verified data from prior tool calls—so that every action is grounded against factual context rather than hallucinated state. When an action is blocked, the system returns structured natural-language feedback explaining why, enabling the agent to retry with corrected parameters. This guided retry mechanism achieves a 58% recovery rate, meaning most blocked actions are self-repaired without human intervention. The system benchmarked a 52% reduction in misalignment on ODCV-Bench across 12 frontier models, demonstrating broad effectiveness across different LLM backends.

It's like having a co-pilot who grabs the steering wheel before you run a red light, then calmly tells you which turn to take instead.

It gives product teams a clear, readable trail of every decision an AI agent made and why—like a flight recorder for robots.

Salus provides a comprehensive observability layer that captures the full trace of every agent run, including each action attempted, the evidence cache state at the time of the action, whether the action was approved or blocked, and the natural-language rationale for the decision. This allows product managers and engineers to audit agent behavior post-hoc, identify patterns in blocked actions, and iteratively refine guardrail policies. Because every decision is logged with its evidence references, teams can quickly distinguish between genuine agent errors (hallucinations, misalignment) and overly aggressive guardrail policies (false positives). This feedback loop accelerates product iteration cycles and builds trust with enterprise customers who require audit trails for compliance. The natural-language format of the logs makes them accessible to non-technical stakeholders, bridging the gap between engineering and business teams in understanding agent reliability.

It's like giving your AI agent a diary that writes itself—except instead of teenage angst, it's a detailed account of every decision and why it didn't send that embarrassing email.

When an AI agent makes a mistake, Salus tells it exactly what went wrong and lets it fix itself—like autocorrect for robot workers.

Salus's guided retry system transforms blocked actions from dead ends into learning opportunities for agents. When the guardrail engine blocks an action, it doesn't simply reject it—it returns a structured, natural-language explanation of the violation, the specific evidence that was missing or contradicted, and suggestions for how the agent can reformulate its action. The agent then retries with corrected parameters, often succeeding on the second or third attempt. This closed-loop self-repair mechanism achieves a 58% recovery rate, meaning that more than half of all blocked actions are resolved autonomously without any human intervention. For operations teams managing large-scale agent deployments (e.g., customer service bots, automated back-office workflows), this dramatically reduces the volume of escalations to human operators, lowers operational costs, and maintains workflow continuity. The system effectively creates a safety net that is both protective and productive—catching errors while keeping the assembly line moving.

It's like a GPS that doesn't just say "recalculating" when you miss a turn—it actually explains what went wrong and suggests a better route before you end up in a lake.