Build tools to help you understand. Build metrics that are proxies for the things you'd care about. Add constraints to the agent's work related to performance, error monitoring, language and architecture choices, review bots (that you tune and eventually, trust). If you don't have good ideas for tools, ask Fable what the best coders and industry experts would look at to establish trust, then feed that into models to build it. Don't sleep on how useful something like a (data-backed) claude artifact can be - i setup one today to track dependency and known vulnerabilties with a burndown chart and (agent) run log, one prompt, it was usable 5 minutes later. A colleague of mine did the same, but as a datadog notebook that also wires in relevant API metrics depending on what code has gone out in the last week, and suggests new longer term metrics to monitor as a result. Etc etc. its a crazy world but this has helped me cope so far.