OpenAI Launches ChatGPT Work, an Agent Built to Finish Multi-Hour Projects
OpenAI introduced ChatGPT Work, a GPT-5.6-powered agent that gathers context across a user's apps and files and turns a stated goal into finished sheets, slides, docs and web apps, staying on a task for hours at a time.
OpenAI launched ChatGPT Work on July 9, a new agent inside ChatGPT built to take a goal, pull context from a user's connected apps and files, and deliver finished output rather than a draft for further editing. The release landed the same day OpenAI took its GPT-5.6 model family public, and ChatGPT Work runs on GPT-5.6 alongside Codex, OpenAI's coding-agent model.
What the agent does
Rather than answering in a single chat turn, ChatGPT Work is designed to gather information across a person's apps and workflows and produce completed materials — spreadsheets, slide decks, documents and simple web apps — by breaking a complex request into smaller steps it executes on its own, staying with a project for hours if the task requires it. The launch also brings Scheduled Tasks to the product: users can have ChatGPT perform an action once, repeat it on a recurring schedule, or trigger it off an event, such as refreshing a recurring meeting agenda whenever a connected Slack channel updates. As part of the release, OpenAI folded its standalone Codex app into the ChatGPT desktop app, so chat, agentic "Work" tasks and coding now live in one client.
Rollout
Availability is staggered by plan and platform. On the ChatGPT desktop app, OpenAI says the agent is available to every plan, including Free. On web and mobile, it started rolling out first to Pro, Enterprise and Edu accounts, with Plus and Business subscribers gaining access over the following days.
Why it matters
ChatGPT Work pushes OpenAI further into the "agent that finishes the job" category already being contested by rivals' coding and computer-use agents, but aimed squarely at general office work — spreadsheets, slides and documents — rather than software engineering alone. Tying the feature to Scheduled Tasks also moves ChatGPT from a tool a person actively prompts toward one that monitors connected systems and acts on triggers without a person re-opening the app each time. That shift raises the same questions now facing every "autonomous" product launch this year: how much oversight a user retains once an agent is running unattended for hours, and how OpenAI will handle mistakes made mid-task across a person's real documents and accounts rather than in a sandboxed test.
Sources
AI-assisted reporting, overseen by the AgentsAI team. Spotted an error? Let us know.
Related agents
More agents news
Nubia to Debut the First OS-Level 'AI Agent' Smartphone at WAIC 2026
ZTE's Nubia brand confirmed its next flagship phone will ship with a system-level AI agent, powered by ByteDance's Doubao AI, that can operate apps and complete tasks like booking flights on a user's behalf — debuting at Shanghai's WAIC 2026 on July 17.
AI Agent Startup Lyzr Used Its Own Agent to Run a $100M Fundraise
Lyzr, a Jersey City startup that builds enterprise AI agents, had its own agent field questions from more than 130 investors and draft memos as it worked toward a $100 million Series B near a $500 million valuation.
Agentic Investing Startup GIM Raises $20M Series A as It Moves to Live Trading
GIM (Grace Investment Machine) closed a $20 million Series A co-led by a US venture firm and Hony Capital, with IDG Capital and Monolith Capital joining, to push its autonomous investment-research agents into live market execution.
Security Researchers Document First Ransomware Attack Run End-to-End by an AI Agent
Sysdig's threat research team disclosed JADEPUFFER, an autonomous LLM agent that broke into an exposed Langflow server, pivoted to a production database, and ran an entire extortion operation — recon through ransom note — without a human operator at the keyboard.