Why AI Coding Agents Need Multiple Personalities to Do Their Best Work

Something curious is happening with AI coding agents. Developers have started telling me they’re asking these coding agents to take on different personas so they behave like more of a team than an individual programmer.

For instance, a developer creating a software product feature will first tell their AI coding product—whether it be Gemini or Codex—to act as a “product manager” that creates a document outlining what the feature is meant to do.

After that, the developer tells them to act as a “spec agent” that translates that document to more specific instructions (known a spec document) outlining what technical work needs to be done; then to act as a “tasks” agent that translates the spec document into specific coding tasks; then a coding agent that executes on those tasks; and, finally, to be a “review” agent that checks the output for mistakes.