Agents that make calls, send email, browse and ship code on their own
Autonomous AI agents on Hard EA run continuously, decide what to do next, and act in the real world without waiting for a prompt at each step. They place real phone calls, send email from their own addresses, message over SMS and WhatsApp, drive a browser, run research and carry out software development tasks. More than 70 production actions are available, including 40 or more for software work.
The runtime is built for long jobs. Agents chain actions in plain natural language rather than rigid JSON schemas, which keeps reasoning quality high while output stays structured. Alloy Mode routes each thought to the model best suited to it, across Claude, GPT, Gemini, Grok and Mistral, so one model's weakness is covered by another. When an agent starts repeating itself, the Loop Buster removes the context that caused the loop and reframes the task instead of blindly retrying. An agent pauses to ask you when only a person can decide, and resumes after a server restart.
The same runtime powers vibe coding: describe what you want in plain language and the agent writes the code, starts a development server, opens the result in a real browser, screenshots it to verify, and iterates until it works. Install the desktop companion app for Windows, Mac or Linux and agents can work directly on your own machine. Everything is also reachable through the developer API, and you can connect your own provider keys.
Open Autonomous AI Agents on Hard EA
What you can do
- Real-world actions: phone calls, email, SMS and WhatsApp, browser automation, research and code
- More than 70 production actions, with 40 or more for software development
- Natural-language action chaining instead of forced JSON function calls
- Alloy Mode routes each step across Claude, GPT, Gemini, Grok and Mistral
- Loop Buster removes the context that causes repetition rather than retrying
- Vibe coding that builds, runs, screenshots and iterates in a real browser
How it works
- Describe the agent's goal and the actions it may take, or pick a ready-made agent from the directory.
- Start it. The agent thinks continuously, chooses actions and executes them, reporting each event as it happens.
- Answer when it pauses to ask you something only a person can decide, such as approving a call.
- Review the results, the transcript and any code or documents it produced.
Frequently asked questions
What actions can an agent take?
Phone calls, email from the agent's own address, SMS and WhatsApp messages, browser automation, web research and software development tasks, from a library of more than 70 production actions.
How is this different from LangChain or AutoGPT?
Agents chain actions with natural-language declarations instead of forced JSON function calls, run on multiple frontier models per task through Alloy Mode, and recover from loops by removing the polluted context rather than retrying.
Can the agents really make phone calls?
Yes. Agents place real autonomous calls with realtime AI voices, and can also email from their own addresses, message over SMS and WhatsApp, and operate a browser.
What is vibe coding?
Describe what you want in plain language and the agent writes the code, starts a development server, opens the result in a real browser, screenshots it to verify and iterates until it works. No coding experience is required.
Can agents work on my own computer?
Yes. The desktop companion app for Windows, Mac and Linux lets agents code and work on your PC. Download it first from the apps page.
Do I have to pay platform token prices?
No. You can connect your own provider API keys and pay those providers directly. The platform's own tokens remain available, with a free daily allowance.