An AI agent builds, tests, and repairs your workflows

In General Input you don't write workflows; an AI agent does. By February 10 the agent could ask you structured questions mid-build, write the workflow, test it in the same sandbox production uses, and repair a failed run starting from its actual error.

How it works

• Questions as forms. Instead of guessing, the agent asks clarifying questions, confirmations, and connection requests as widgets in chat, then resumes exactly where it stopped once you answer.
• Tested against real services. Test runs use your connected accounts, and the agent must read the output before it reports back. Its instructions forbid mock data and placeholder fallbacks that make a run look successful without doing the job.
• Fix with AI. Any failed execution has a Fix with AI action that opens a new conversation seeded with that run's state and logs.
• Long conversations keep going. When a conversation nears the model's context limit, it is summarized into the request, current state, pending work, and key decisions, and the agent continues from there.

Also shipped

• Every agent edit saves a version you can restore.
• A model picker in chat, and a running total of what building a workflow has cost.