Enforce plain-English safety policies on AI agent decisions at inference time.
npx skills add https://github.com/zentropi-ai/reflexes --skill reflexes
Please help me install this Agent Skill. Skill: reflexes Source: https://github.com/zentropi-ai/reflexes/tree/main/skill Command: npx skills add https://github.com/zentropi-ai/reflexes --skill reflexes
AI agents often drift, executing destructive commands or ignoring instructions. This Skill provides a safety net that checks agent decisions against plain-English policies in real-time, catching mistakes before they occur.
Ask your AI agent to install the reflexes skill and follow the numbered steps in the INSTALL.md file to wire the safety harness into your host environment.
High-intent search queries and answers about installing and using this skill.
Agent self-correction works by denying actions that violate plain-English policies and feeding the policy text back to the AI agent. This real-time inference-time feedback loop enables the agent to correct its own course and retry the action safely.
CLI hosts and Python agent frameworks are supported through a configurable hook and middleware architecture. This allows the inference-time safety harness to validate agent actions across different agentic workflows and host environments.
The best way to stop destructive commands is implementing real-time inference-time guardrails that validate agent tool calls against plain-English safety policies. The action is denied before execution if it violates the defined compliance rules.
You need uv, yaml, and pydantic dependencies to set up AI agent policy enforcement. These tools support the configurable hook and middleware architecture that wires the safety harness into your host environment.
RaudbjornGenerate automated migration scripts for evolving Pydantic configuration schemas.
selfagencyEnforce allowlists, blocklists, and content-policy checks on AI agent tool usage.
woody1234567Add human-in-the-loop approval and custom middleware to LangChain agent tool calls.
TserenTserenovEvaluate proposed architectures with a seven-metric EMOGSSB profile and blocking questions.
