ChatGPT Tasks: Scheduled AI Agents Inside ChatGPT

OpenAI quietly released one of the most important features ChatGPT has received in months: the ability to schedule reminders and recurring tasks. On the surface, it looks like a simple addition. Set a reminder, get a notification. But underneath, this is OpenAI laying the groundwork for AI agents that take action on your behalf at specific times, without you being in the conversation.

The feature shipped as part of GPT-4o with scheduled tasks. You describe what you want in natural language, set a time or interval, and ChatGPT handles the rest. When the task fires, you get a notification. Click it, and you open a conversation thread with the results.

How Scheduled Tasks Work

The setup is straightforward. You type something like "send me the latest AI news at 8:00 AM" and ChatGPT creates a recurring task. When 8:00 AM arrives, the model runs a search query using its web search capabilities, gathers the latest results, and assembles them into a conversation. You receive a notification, click through, and see a curated AI news briefing with sources.

The scheduling interface uses natural language entirely. You do not need to configure cron expressions or fill out forms. Say "8:00 AM and 5:00 PM" and it sets two daily triggers. Say "every 3 months" and it creates a quarterly reminder. The model interprets your intent and translates it into a schedule.

Once a task is created, you can manage it through the tasks panel. Click the three dots menu, and you see all scheduled and completed tasks. Each one can be edited, paused, or deleted. The edit modal lets you change the task name, instructions, and schedule. You can toggle whether it repeats and adjust the timing.

Practical Use Cases

Daily News Briefings

The most obvious application is automated research at regular intervals. Ask ChatGPT to send you industry news every morning, and it acts as a personalized news aggregator. Because it uses the same search infrastructure as ChatGPT's web search, the results include sources and are formatted the same way as a manual search query.

This is more useful than a traditional RSS feed or news app because you can customize the scope with natural language. "Send me AI infrastructure news, but only funding rounds over $50M and new model releases" is a perfectly valid task description. The model will filter and curate based on your specific criteria.

Weather Updates

A simple but practical example: "Send me the weather every morning at 7 AM." The model pulls current weather data and delivers a brief summary. It is not going to replace a dedicated weather app, but it demonstrates the pattern. You can ask for any information that ChatGPT can retrieve through search, delivered on a schedule you define.

Fitness and Health

One of the more interesting applications is personalized workout planning. Ask ChatGPT to "send me a workout plan every day at 8 PM using dumbbells and a stationary bike" and you get a unique, varied plan each day. Because ChatGPT maintains context about your preferences across the conversation, the plans can build on each other over time.

This starts to challenge dedicated fitness apps. The advantage of a general-purpose AI over a specialized app is flexibility. You can dump in information mid-workout - "I just did 3 sets of 12 at 30 pounds on overhead press" - and the model adjusts future recommendations accordingly. You could even use voice mode to have it coach you through exercises in real time.

Creative Content Generation

You can schedule creative outputs too. "Generate a children's bedtime story about dragons every night at 9 PM" produces a new story with an AI-generated image each evening. Add instructions like "also write a script for the story" and you get both visual and narrative content on a recurring basis.

With OpenAI's audio generation capabilities evolving, it is easy to imagine this extending to audio stories, personalized podcasts, or daily creative writing prompts delivered at whatever time suits your routine.

Price Monitoring and Research

One of the most forward-looking examples from the announcement: "Research the best price on furnace filters every 3 months and have one delivered to my door." This does not work today. ChatGPT can research the prices, but it cannot execute a purchase. However, the task infrastructure is clearly designed to support this kind of workflow once the agent capabilities expand.

The model interpreted this request as: find a furnace filter, search for the best price, and notify when there is a good deal. The notification step works now. The purchase step is what is coming next.

Get the weekly deep dive

Tutorials on Claude Code, AI agents, and dev tools - delivered free every week.

Where This Is Heading

Sam Altman published a blog post around the same time stating that 2025 could see the first AI agents "join the workforce and materially change the output of companies." The tasks feature is the foundation for that vision.

Today, the tasks are observe-and-notify. The model watches for something, gathers information, and tells you about it. The next step is observe-and-act. The model watches for something, gathers information, and takes action on your behalf. The infrastructure for scheduling, notifications, and task management is already built. What remains is expanding the model's ability to interact with external services.

Consider the progression:

Search and report (available now) - "Tell me the AI news every morning"
Monitor and alert (available now) - "Let me know when GPU prices drop below $X"
Monitor and act (coming) - "When GPU prices drop below $X, buy one from the cheapest vendor"
Multi-step workflows (coming) - "Every quarter, research the best deals on home supplies, compare against my past purchases, and order anything that is a better deal than what I paid last time"

Each step requires the model to have more agency - more ability to interact with the outside world. OpenAI has already shipped Operator (their web browsing agent) and Deep Research (their research agent). Tasks is the scheduling layer that connects these capabilities to recurring workflows.

The Notification System

At launch, the notification system had some limitations. Tasks delivered results through the ChatGPT conversation interface. You receive an email notification with a preview and a link back to the conversation thread. The desktop and mobile apps were installed during testing, but push notifications were not consistently firing.

This is expected for a beta feature. Push notifications on mobile are essential for this to feel like a true personal assistant rather than an email subscription service. The infrastructure is clearly designed for it, and consistent push notification support would make a significant difference in the day-to-day utility of scheduled tasks.

Tasks vs. Dedicated Apps

The emergence of scheduled AI tasks raises an interesting question about the future of specialized applications. Consider fitness apps, news aggregators, weather apps, recipe planners, and budget trackers. Each of these exists because they solve a specific problem with a purpose-built interface. But a general-purpose AI that can take instructions in natural language and deliver results on a schedule competes with all of them simultaneously.

The advantage of specialized apps is their refined UI, hardware integration (like Apple Health syncing for fitness), and deep domain knowledge baked into the product. The advantage of ChatGPT tasks is flexibility. You can combine any number of capabilities into a single workflow without switching between apps. "Check the weather, then suggest an outfit, then add my commute time to my calendar" is one task description that would require three separate apps otherwise.

In practice, specialized apps will not disappear. They offer things that a chat-based interface cannot - real-time heart rate monitoring, interactive maps, collaborative editing. But the simple, information-retrieval-and-action category of apps faces genuine disruption from AI agents that can do the same thing through natural language.

What Developers Should Pay Attention To

For developers building products on top of OpenAI's platform, the tasks feature signals that the API will eventually support scheduled and recurring agent interactions. This opens up new application patterns:

Monitoring services that use LLM reasoning to interpret unstructured data (news, social media, forum posts) and deliver structured alerts
Workflow automation where the scheduling and routing logic is described in natural language rather than configured through a visual builder
Personal assistants that maintain long-running context across scheduled interactions, building up a knowledge base about the user's preferences and history over time

The key technical detail is that each task fires within a conversation thread. This means the model has access to the full conversation history when executing a scheduled task. Over time, this creates a rich context about what the user has asked for, what results have been delivered, and how preferences have evolved. That context is what separates a scheduled search query from a genuine personal assistant.

Current Limitations

The feature is still in beta, and several limitations are worth noting:

No action execution - Tasks can search and report but cannot take actions like making purchases, sending emails to third parties, or modifying external services.
Notification reliability - Push notifications are inconsistent across platforms. Email notifications work but add friction to the experience.
No integration layer - Tasks operate within ChatGPT's existing capabilities (search, code execution, image generation). There is no way to connect them to external APIs, databases, or services yet.
Rate limits - Like all ChatGPT features, tasks are subject to rate limits based on your subscription tier.

These are solvable limitations, and most of them are likely on OpenAI's roadmap. The foundation is solid. The question is how quickly the execution capabilities expand to match the scheduling infrastructure that is already in place.