Coming soon Emerging

Agents that finish multi-step tasks

AI that plans and carries out longer tasks for you. Early versions exist; reliability is still improving.

Forward-looking: this describes work in progress, not a finished product. Nothing here is a promised date or release.

What is it?

Today most AI tools answer one request at a time. AI agents aim to go further: you describe a goal, and the agent plans steps, uses tools such as a browser, files or apps, checks its progress and keeps going until the task is done or it needs your input.

Early versions already exist, for example agents that browse websites, fill in forms or make changes across a code project. Many researchers and companies expect them to handle longer and more varied tasks, but how fast this happens, and how reliable and safe it becomes, is uncertain.

Which models use it?

OpenAI, Anthropic and Google have released early agent features that can operate a browser or computer for some tasks.
Coding agents such as Claude Code, GitHub Copilot's agent features and OpenAI Codex can work through multi-step programming tasks.
Open-source frameworks such as LangGraph and AutoGen help developers build their own agents.
The Model Context Protocol (MCP) is an open standard that lets agents connect to tools and data.

How to use it

Start with low-risk tasks, such as research summaries or drafting, before trusting agents with important actions.
Describe the goal, limits and what 'done' looks like clearly.
Give the agent only the access it needs, and keep sensitive accounts disconnected unless required.
Ask the agent to show its plan first and to pause before irreversible actions like payments or sending messages.
Review the result and the steps it took, not just the final message.
Expect mistakes and keep a way to undo changes.

Benefits

Could save time on repetitive multi-step work such as gathering information or filling forms.
Lets people describe goals in plain language instead of learning many separate tools.
Can work through long tasks while you focus on decisions.

Limits and risks

Current agents can get stuck, misunderstand goals or take wrong actions.
They are vulnerable to prompt injection from web pages and documents they read.
Giving an agent access to accounts and money creates real risks.
Timelines and capabilities are uncertain; claims about future agents should be treated with caution.

Prompt template

Copy it, replace the parts in [brackets] and paste it into any AI chat.

Act as a careful assistant planning a multi-step task. Goal: [what I want done]. Tools or information you can use: [list]. Constraints: [budget, time, things not to do]. First, write a numbered plan and list any questions. Do not take any irreversible action (payments, sending, deleting) without asking me. After each step, report what you did and what is next.