Not a bug report — noticed some real design convergence, curious if there's a conversation worth having
We're Tulip (tulip-agents on PyPI, Apache-2.0, pre-GA — flagging that plainly). We build an admission-gate + audit-trail layer for AI agents taking real actions: policy-checked per call (allow / require-human / deny), a tamper-evident hash-chained record of every decision.
While looking at this repo for an unrelated reason, we read security/approval_handler.py and security/audit_logger.py — the Security Envelopes (RO/RW/RESEARCH/DEPLOY), the three approval modes, and specifically "tamper-evident JSONL logs with hash chaining." That's independently, convergently, almost exactly the same architecture we've built as tulip.control.admit() + AuditTrail. Genuinely interesting to see the same shape arrived at independently for the same problem — usually a decent signal the shape is right.
Not proposing you rip out something that clearly already works well for this specific tool — you've built real, working infrastructure. The actual reason for opening this: is there a broader conversation worth having about the same primitive across more of what Snowflake's agents (Cortex Agents, etc.) do? Happy to share what we've learned building this pattern across a few different real integrations (payments APIs, communications APIs, MCP servers) if that's useful, or just to compare notes. No ask beyond that — genuinely curious, not pitching a replacement for something that isn't broken.
Not a bug report — noticed some real design convergence, curious if there's a conversation worth having
We're Tulip (
tulip-agentson PyPI, Apache-2.0, pre-GA — flagging that plainly). We build an admission-gate + audit-trail layer for AI agents taking real actions: policy-checked per call (allow / require-human / deny), a tamper-evident hash-chained record of every decision.While looking at this repo for an unrelated reason, we read
security/approval_handler.pyandsecurity/audit_logger.py— the Security Envelopes (RO/RW/RESEARCH/DEPLOY), the three approval modes, and specifically "tamper-evident JSONL logs with hash chaining." That's independently, convergently, almost exactly the same architecture we've built astulip.control.admit()+AuditTrail. Genuinely interesting to see the same shape arrived at independently for the same problem — usually a decent signal the shape is right.Not proposing you rip out something that clearly already works well for this specific tool — you've built real, working infrastructure. The actual reason for opening this: is there a broader conversation worth having about the same primitive across more of what Snowflake's agents (Cortex Agents, etc.) do? Happy to share what we've learned building this pattern across a few different real integrations (payments APIs, communications APIs, MCP servers) if that's useful, or just to compare notes. No ask beyond that — genuinely curious, not pitching a replacement for something that isn't broken.