How ChatGPT Work's agent architecture actually works, reconstructed

Latent Space published an external reconstruction of how ChatGPT Work, described as an agent built for a billion users, handles memory, proactivity, scheduling, browser use, plugins, skills, and tools.

Why it matters: As agent products from OpenAI, Anthropic, and others converge on similar capabilities like memory and scheduling, independent architecture breakdowns like this are one of the few ways outsiders can understand how these systems are actually engineered, absent official documentation.