A newer class of tools writes policy directly over agent traces, tool names and arguments, or enforces it at an MCP gateway. Coding agents also ship their own rule files for command-level allow, prompt and forbid decisions.
Read in this order
- OfficialRules (OpenAI Codex docs)Start here: command-level allow, prompt and forbid rules in a shipping agent.
- Open sourceInvariant guardrails (Invariant Labs)Policy written over agent traces and tool calls, with runnable examples.
- PaperProgent: Securing AI Agents with Privilege ControlProgent's privilege control approach, for the research behind tool-level policy.