How are you handling agent security in production? (Identity, permissions, kill switch)

LangSmith Product Help
1

Hey everyone :waving_hand:

I’ve been deploying LangChain agents in a fintech context and ran into a serious gap — there’s no built-in way to:

  1. Verify which agent is executing a tool call (cryptographically, not just via prompt)

  2. Enforce hard limits like “$500 max per transfer” at the protocol level

  3. Instantly kill a compromised agent across all tools

I ended up building an open-source protocol for this called AIP — gives each agent an Ed25519 identity, signed intent envelopes, and a real-time kill switch. All sub-millisecond, no network calls.

Curious how others are handling this? Are you relying on prompt-level guardrails, API key scoping, or something else?

Would love to hear what’s working (or not) in your production setups.