Steering Agents is an Unsolved Problem, says OpenAI’s Alexander Embiricos

When OpenAI introduced Codex in August 2021 it was framed as a model that translates natural language into code. Today’s Codex has evolved: product lead Alexander Embiricos tells the AI+16z podcast that the product is meant to be “an agent working remotely”. A cloud teammate that can perform tasks in sandboxed environments and propose pull requests for review. Embiricos describes a development arc from small, single-file autocompletions to multi-file reasoning and background execution. “You can give Codex a description, and it will generate a pull request on its own,” he said, noting the product’s cloud-agent design and the safety tradeoffs that informed it.  OpenAI documents that each Codex task runs in its own sandboxed environment and that agents operate in repo-scop