3/25/2026 at 11:37:29 PM
The staged autonomy pattern ("trust is earnable") maps directly to what we built with protect-mcp — shadow mode first (log everything, block nothing), then enforce when you've seen enough data to trust the policies.For the prompt injection concern: protect-mcp wraps MCP tool calls with per-tool policies. Even if the agent gets injected, it can't call tools outside the policy. Every decision is optionally Ed25519-signed and verifiable offline.
npmjs.com/package/protect-mcp
by tomjwxf
3/26/2026 at 9:32:16 AM
hm, interesting! I like, what I've done is that each step in the process is one agent. One agent get's one task, and only the tool to do it.by Mrakermo
3/26/2026 at 11:49:46 AM
[dead]by tomjwxf