Morning Edition · Friday, July 10, 2026Published at 1:31 AM EDT · New York
OpenAI Launches ChatGPT Work, an Agent That Runs Multi-Hour Tasks Across Apps
The product combines Codex into a single desktop app and adds a plugin directory spanning Google Drive, Slack, Salesforce, and GitHub.
Alongside GPT-5.6, OpenAI launched ChatGPT Work, an agent that can take action across a user's apps and files, run on a project for hours, and return finished work such as a spreadsheet, slide deck, or website. It is powered by GPT-5.6 and the Codex coding engine, and OpenAI is merging the standalone Codex app into a single ChatGPT desktop client for Mac and Windows.
The most consequential piece for engineers is the plugin surface. ChatGPT Work ships with a unified directory of third-party integrations that at launch includes Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, GitHub, and Dropbox, letting the agent read from and act on enterprise data stores directly. The AI Post channel described it plainly as an "AI coworker" that works on projects for hours rather than answering single prompts.
The claim to weigh is durability. An agent that stays with a project for hours is only as useful as its error recovery and its handling of stale or conflicting state across a dozen connected systems, and that is exactly where autonomous agents have historically degraded. OpenAI is rolling access to Pro, Enterprise, and Edu tiers first on web and mobile, while the desktop app is available on every plan, including the free tier.
What this means
OpenAI is moving from selling a model to selling an agent that operates across a company's existing software, which makes the plugin directory its main competitive advantage. The exposed parties are vertical agent startups and Microsoft Copilot itself, since a general agent wired into Salesforce, Slack, and GitHub competes directly with task-specific tools through distribution rather than superior reasoning. Whether this holds depends on whether multi-hour runs finish correctly or fail partway without warning.
What to watch
- Documented completion and error rates on long-running ChatGPT Work tasks, which separate a genuine autonomy advance from a demonstration that breaks past the first hour.
- How enterprises govern an agent with write access to Salesforce, Gmail, and GitHub, since one wrong tool call in a connected system is a data or production incident.
Observations to monitor, not financial advice.
Synthesized from: OpenAI · Polylog editors
Part of a tracked trend
Agentic AI Moves Into Enterprise and Government Workflows
Over the next 3-9 months, AI agents move from demos into real enterprise and public-sector workflows, with deployment success tied to domain and task understanding more than raw model capability.
More from this edition
- OpenAI Ships GPT-5.6 Family to the Public and Claims the Top Coding Benchmark
- US Commerce Department Clears GPT-5.6 for Full Public Release
- Developers Push China's Open-Weight GLM 5.2 Onto Modest Consumer Hardware
- Ollama Raises $65 Million as Open-Model Tooling Reaches 8.9 Million Developers
- Meta's Brain2Qwerty v2 Decodes Typed Sentences From Brain Scans at 61 Percent Word Accuracy
- Survey Charts LLM Theorem Provers Moving From Solving Problems to Doing Research Math
- Theory Paper Asks When Reflection-Driven Reasoning Actually Helps LLMs
- DeepSearch-World Trains Tool-Use Agents to Improve From Their Own Experience
- AgentLens Argues Code-Agent Benchmarks Should Grade the Whole Trajectory, Not a Pass/Fail Bit
- EU Parliament's Effort to Kill 'Chat Control' Message Scanning Falls Short
- ByteDance Releases Seedream 5.0 Pro, an Image Model That Outputs Editable Layers