Unpacking the Agentic Architecture of ChatGPT Work
Latent Space has published an in-depth external reconstruction and analysis of the system architecture behind ChatGPT Work's agentic capabilities. The teardown details how components like memory, proactivity, scheduling, browser use, and tool integration function together. Understanding the architectural design of a production-grade agent like ChatGPT Work provides crucial blueprints for engineers building scalable, real-world LLM agents. It highlights how AI is transitioning from simple reactive chat interfaces to proactive, multi-tool systems. The analysis covers how ChatGPT Work manages persistent memory across sessions, schedules proactive tasks, and coordinates external tools like web browsers and plugins. It addresses the engineering trade-offs required to scale these agentic workflows to a massive user base.
## BACKGROUND
Traditional LLM applications operate on a simple reactive "Agent Loop" where the model responds to user prompts one at a time. Modern agentic architectures are moving toward structured graphs and advanced scheduling systems to handle proactive tasks, manage persistent memory across sessions, and execute complex tool-calling workflows.