Unpacking ChatGPT Work: OpenAI's Cloud Agent for a Billion Users
Latent Space · rss · 2026-08-05
Latent Space provides an in-depth breakdown of OpenAI's newly released agent product, ChatGPT Work. Reaching 10 million users within three weeks, the product is seen as a preview of how ChatGPT will eventually merge chat and work modes for its massive user base.
Core Architecture & Execution
- Cloud MicroVMs: Tasks run in isolated microVMs (Pro accounts get 8 CPUs / 20GB RAM) equipped with a managed Chrome service for agent operations.
- Codex Foundation: Inherits Codex's models, sub-agents, and browser capabilities to grind on long tasks, but strips away coding-specific UI to appeal to non-developers.
- Local and Cloud Sync: Desktop apps support both local file operations and cloud execution, trackable via mobile, though local and cloud tasks don't yet interoperate.
Memory & Persistence Design
- Workspace Persistence: Each task gets an isolated /workspace/scratch directory where the agent can create folders, write scripts, and maintain databases, with state synced to the cloud.
- Context Isolation: Cross-task context isn't shared via raw directories. Instead, it relies on a Personal Context tool to fetch summaries, and a central Library for file management.
- External Memory: User profile memory is maintained asynchronously and injected into tasks; the agent cannot directly modify the underlying memory files.
More from coding & agent
- Fixing Node.js Bugs with DeepSeek Costs Just $0.034 in Real-World Test — film_girl · 2026-08-05
- Cloudflare Open-Sources 'computer', Shifting Agent Compute to Per-Process Model — dotey · 2026-08-05
- A 9-Layer Stack Guide for Building Production AI Agents — MaryamMiradi · 2026-08-05
- Open-Source Channels SDK: Connect Any AI Agent to Slack and Teams — _jaydeepkarale · 2026-08-05
- "Verification Sharding": Distributing Output Validation Across Massive Agent Swarms — curious_vii · 2026-08-05
- Building Software on the Eve of ASI: Six Core Opportunities — pzakin · 2026-08-05