Shadow Run: let your agent break a clone first

AI agents are getting the keys to production. Shadow Run is how a team says yes to that safely: the agent never touches prod directly. Every risky change it wants runs first on a disposable copy of production, and what the change really did, not what it says it will do, is what a person approves. Today, on Vultr: the agent gets no credential. Each SQL or shell plan is frozen, hashed and rehearsed on a full clone of prod in a throwaway microVM on a Vultr VX1, cut off from the network. Postgres measures the blast radius in rows, cascades and dollars, frozen guardrails decide, a different person approves those exact hashes, and the same plan runs on prod in one transaction that rolls back if anything it measures differs. Every model call goes to Vultr Serverless Inference. It works. On 23 held-out risky tasks, even with every plan approved unread, harm fell from 10 to 2.67 (p = 0.0078 over tasks). 7 of 7 live attacks were contained and 6 of 6 real incident commands stopped before prod. It also found something new: told which rule it broke, an agent can get around it while every guardrail still passes. The checks this led to cut harm on outside-benchmark tasks from 5.33 to 2. Where it goes: the best version of Shadow Run is the default gate between every coding agent and every production system. Next come thin clones and clones from Vultr Managed PostgreSQL's own backups, so a rehearsal stays fast past a gigabyte; an auto-apply list that learns from what a team approves, tested against its own attacker first; one-click deployment on a team's own Vultr account; and the free pull request Action on GitHub Marketplace as the way in. Who pays: the engineer who owns production Postgres at a team using coding agents. It's open source and free to self-host on Vultr; hosted, it's $500 a month per protected database, at about a 70% infrastructure margin on one database. Race the agent: https://144-202-92-64.sslip.io/race