OpenAI Launches ChatGPT Agent for Multi-step Tasks

OpenAI has flipped the switch on a new ChatGPT “agent” mode that goes far beyond answering questions starting last week, Pro, Plus and Team subscribers can ask the chatbot to plan and execute multi‑step tasks—everything from ordering a full wedding outfit online to booking fittings, tracking the deliveries and dropping reminders straight into Google Calendar. Under the hood, the agent spins up a private “virtual computer” equipped with its own browser, file system and productivity toolbox, then works through the steps much like a junior assistant: it opens e‑commerce sites, fills forms, checks weather forecasts, compares prices, schedules appointments and reports back once the job is done.

The launch merges two earlier experiments, Operator (an autonomous web‑automation bot introduced in January) and Deep Research (a multi‑hour web‑scraping assistant rolled out in February), folding both into a single toggle labelled “agent mode” inside ChatGPT. OpenAI’s new system card says the tool chains GPT‑4.1 calls with deterministic “skills,” such as read‑site, click‑button or send‑email, and requests an explicit user confirmation before any high‑stakes action like making a payment. CEO Sam Altman admits the agent is still experimental, warning users not to rely on it for tasks such as wire transfers or confidential legal filings; safeguards include output attribution, rate limits and a refusal to handle sensitive credentials.

Early testers can already hook the agent into Gmail, GitHub and Slack, with Shopify checkout integration in preview—a hint that OpenAI plans to take a commission on purchases placed inside ChatGPT much the way Amazon collects marketplace fees. A separate Financial Times leak says the company is building an end‑to‑end payment rail so that merchants can embed one‑click buy buttons directly in chat responses, broadening revenue streams beyond the US $20‑per‑month subscription.

OpenAI’s move lands in the thick of an emerging agentic‑AI land‑grab: Microsoft has Copilot plug‑ins that automate Office workflows, Google’s Gemini is rolling out “Nano” agents inside Android, and Salesforce, Anthropic and Perplexity are all pitching agents that draft emails or crunch spreadsheets. By shipping a consumer‑ready solution first—and piggybacking on ChatGPT’s roughly 300 million‑strong user base—OpenAI hopes to cement its position as the place where mainstream users experience autonomous AI for the first time.

Not everyone is cheering. Privacy advocates point out that letting a single model fetch emails, browse shopping carts and schedule calendars concentrates enormous sensitive data in one place, making any breach or jailbreak far more damaging. Meredith Whittaker, president of Signal, called the agent “an over‑permissioned honeypot” and urged regulators to scrutinise how granular OpenAI’s consent prompts really are. OpenAI counters that each tool call is logged, sandboxed and subject to automated policy checks, and it published a 41‑page system card detailing red‑team results, hallucination rates and disallowed content categories.

For power users, the practical upside is time: the demo Reuters viewed showed the agent planning a three‑day bachelor‑party itinerary—flights, hotel blocks, restaurant reservations and an Airbnb split payment—in 12 minutes, a task the company says took human assistants “most of an afternoon.” Developers will get an API hook later this quarter, enabling SaaS apps to hand off repetitive browser work to ChatGPT the same way they once wrote Zapier integrations. In the meantime, subscribers can find the new feature live today under ChatGPT’s mode menu; switch it on, describe your project, and watch the agent open its own mini‑desktop to get the job done—proof that the future of chatbots is less about talk and more about getting real‑world work off your plate.