Quarantine

Quarantine is a self-improving harness for an IT-operations agent. Every tool call goes through one MCP gateway that enforces a policy stored in MongoDB Atlas. The policy starts empty. After a failed check, an Evolver agent writes the next policy version, which goes live only if it is at least as strict and passes an eval on recorded traces. In the demo, a helpdesk ticket carries one false sentence: corporate DNS resolvers moved to 203.0.113.66. Under the empty policy v0 the agent saves it to shared memory, the next session builds server web-07 with it, and the server fails its health check. The Evolver then writes policy v1: reading a ticket taints the run, and memory writes from tainted runs are held. A Guardian agent checks each held write against trusted memory and flags contradictions, and a person releases or purges it in a Streamlit console. A second failure, a DNS record change planted in a ticket, produces policy v2, which keeps v1 and denies such changes from tainted runs. Nobody edits the policy by hand. Measured: in a replay of 20 tickets, 10 of them poisoned, 9 of 10 poisoned settings reached a server under v0 and 0 of 10 under the learned v1. The cost is that all 10 harmless notes were held for review. On 32 recorded traces, v0 blocks 0 of 12 attacks, v1 10 of 12 and v2 12 of 12, with all 10 clean runs unaffected. Real: all data in the Atlas sandbox cluster, the gateway and its enforcement, the worker, Evolver and Guardian model calls, and the learned policies, which are written at run time. Mocked: servers, the monitor and Route 53 are documents; tickets are seed data.