← sources

Agent guardrails on the traffic or on the tool call?

source: reddit (reddit:AI_Agents)published 2026-11-02 19:41ingested 2026-11-03 02:36

retained text

So as I read, there are several new agent guardrails that went by this week. One is a proxy that applies policy to AI traffic and redacts keys. One is a browser extension that stops secrets before they reach a chat model. There is cryptographic identity for agents calling MCP servers, and GitHub Actions that record what an agent did. They all sit at different layers, and I think the layer matters more than the feature list. So I wanted to understand what is best practices around guardrails.

A proxy on model traffic sees text going out and text coming back. It can redact a key or block a destination. What it does not see well is the moment the agent decides to do something irreversible, like deleting a directory or sending an email. That decision happens in the harness, not on the wire.

I built my own runtime on the Claude Agent SDK, Mainsheet, around the other choice. A gate in code checks every tool call before it runs. It checks argument patterns, deny patterns, budgets and human approval. Once a tool marked as returning untrusted output has run, tools marked irreversible are refused for the rest of the session. Each decision becomes a signed receipt in a hash chain.  
That layer has its own gaps, and I would rather name them. The taint is one flag for the whole session, not a label on each piece of context, so it is coarse. A receipt is also only as trustworthy as the place the signing key lives, and I have not documented that yet.

So here is my question for people running agents outside a demo. Where did you put the enforcement? Has a proxy ever caught something a tool-call gate would have missed, or the other way round?

document fetched 2026-11-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xm7988/agent_guardrails_on_the_traffic_or_on_the_tool