Salus - Block Incorrect Agent Actions

What

Salus is a runtime safety layer for AI agents that validates tool calls before they execute. It is designed to prevent incorrect or unsupported agent actions by intercepting calls, checking them against policy and evidence requirements, and blocking them when they fail validation.

The product appears aimed at teams building agentic workflows with external tools, especially where mistakes can trigger costly downstream actions such as refunds or other commits. Based on the page, Salus is positioned as an agent guardrail and evaluation product for development and production use, with lightweight code-level integration.

Features

• Runtime guardrails — Intercepts each tool call at execution time and blocks actions that violate policy or are not grounded in evidence before they run.
• Self-repair feedback — Returns structured error information when an action is blocked so the agent can retry with a corrected action.
• Execution visibility — Provides real-time streams of agent interactions, including traces, token usage, and latency breakdowns for debugging and oversight.
• Pre-deployment evals — Generates large sets of adversarial and realistic scenarios to test tool-call correctness before release.
• Minimal tool-level integration — Uses a decorator-based approach for protecting tool functions, which can reduce implementation effort in existing agent code.
• Framework compatibility — The page states support for OpenAI, Anthropic, LangChain, LangGraph, and CrewAI, which suggests Salus can fit into common agent development stacks.

Helpful Tips

• Focus evaluation on high-risk tool calls first, especially actions that commit changes, move money, modify records, or affect customers.
• Use evidence-producing steps and explicit dependency rules in the workflow so validation can distinguish between informed actions and unsupported ones.
• Review blocked-action traces and structured failure reasons regularly; they can expose prompt design issues, weak tool schemas, or missing retrieval steps.
• Test both adversarial and routine scenarios before deployment, since agent failures often come from ordinary edge cases rather than only malicious inputs.
• Confirm how Salus fits into your existing observability and incident processes, because the page shows visibility features but does not describe broader operational integrations.