Imbue

Imbue's Sculptor sits somewhere between a coding assistant and a full autonomous agent. You give it a task in natural language, it spins up a secure sandbox, writes code, runs tests, and iterates until it's done or hits a wall. The sandboxing is the headline feature: every agent session gets a fresh Ubuntu container with network access, so you can let it install dependencies, run scripts, and poke APIs without worrying it'll trash your local environment. I tested it on a Python ETL script that needed to pull data from a REST API, transform it, and write to SQLite. Sculptor handled the boilerplate competently, installed requests and pandas without prompting, and caught a type error in its own code after the first test run. The UI shows you a live terminal feed and file diffs, which is reassuring when you're letting an agent loose. It took three iterations to get the schema right, but I didn't have to intervene. Where it stumbles: reasoning about ambiguous requirements. If your spec has gaps, Sculptor will make assumptions and run with them. I asked it to "add error handling" and it wrapped everything in try-except blocks that silently swallowed exceptions, which is technically correct but useless. You need to be specific, or you'll spend time undoing its choices. Vet, their open-source review tool, is a nice companion. It logs every action the agent takes and lets you audit decisions after the fact. Useful if you're paranoid about what code is running in that sandbox, or if you want to understand why the agent went down a particular path. Compared to Devin or Cursor's agent mode, Sculptor feels more research-y and less production-ready. The waitlist and freemium model suggest they're still figuring out pricing and scale. If you're an early adopter who wants to experiment with autonomous coding in a controlled environment, it's worth the wait. If you need something you can rely on for client work tomorrow, stick with Cursor or Copilot Workspace.