Essentially I'm building a sandbox that I can let an agent loose in, I'm building it to use codex, claude code, or AMP code. I have a an offline version of the retell docs for the agent to reference, and a master prompt that tells the agent what I need and where to find docs, and how to use the test suite. It generates a directed graph of the full agent flow (later shoved into retell through api calls), validates everything through a zod schema, does some algorithmic checks on the flow to check for unreachable nodes, loops, etc. and finally runs tests on it through a local state machine to ensure proper data collection, and the right endpoints get reached. Finally everything gets put into retell through the SDK with auto generated simulation scenarios. It works well in all aspects except for the node prompts. The high level prompt is great, but the individual node prompts it comes up aren't great for now. It does great with creating custom tools and putting them into the SDK though. The thought is that the agent iterates against the local tools I've given it until it has something that is guaranteed to be valid, with tool calls connected to the necessary api's with valid keys to user provides, and a test suite the user can run in retell as a final check. The goal is to drop the barrier to entry for non-technical folks to build these agents, and to build them faster. I'd be happy to bounce some ideas off of you