Marketplace onboarding, canvas loops, and SQL over spreadsheets

This release makes MagOneAI easier to buy and more capable to build with. You can buy it on a cloud marketplace and start immediately. Workflows can loop back to improve a result, agents answer spreadsheet questions with exact figures, and agents can stop and ask a person for help.

Highlights

Buy on a cloud marketplace and start immediately

MagOneAI can now be bought through a public cloud marketplace and used right away, with no setup call needed. Buyers onboard themselves: they create a new organization or attach the purchase to one they already own. The usage allowance for the plan they bought is applied automatically.

AWS Marketplace is live today. Azure and Google Cloud are in preview. Superadmins get a Marketplace page to switch each marketplace on and map each plan to what it includes.

Why it matters: Your organization can buy MagOneAI through the cloud account it already uses, and get started the same day.

Learn more: Cloud Marketplace

Loops on the canvas

A Condition node chooses which path a workflow takes next. It can now point back to an earlier step, so part of a workflow repeats until the result is good enough. For example: draft a reply, check it, and redraft if the check fails. You set a limit on how many times it can loop, and a branch to take when the limit is reached, so a loop can never run forever.

Why it matters: "Try again until it's right" no longer means copying the same steps over and over on the canvas.

Learn more: Condition node

Exact answers from spreadsheets

Knowledge bases built from CSV and Excel files are now queried with real database queries (SQL) instead of the AI model reading the numbers. Totals, averages and counts are exact, not the model's best estimate.

Why it matters: You can trust the numbers an agent gives you from your spreadsheets.

Learn more: Knowledge bases and RAG

Agents can ask a person

An agent can now pause during a run, ask a named person a question, and carry on once it gets the answer. You can ask up to 50 people at once, and the first reply wins.

Why it matters: When an agent lacks information or needs a judgment call, it can ask someone instead of guessing.

Learn more: Human task node

Improvements

• Prompt caching per agent. Many AI providers can reuse a repeated part of a prompt at a lower price, which is called prompt caching. Each agent now has a setting for how long cached content is kept. Cost tracking takes cached tokens into account, and a Cache Hit tile shows what you're saving. See Analytics.

Bug fixes

• Notifications are now sent. Alerts about tool server status, schedule changes, failed runs and credential changes now arrive by email and in the in-app notifications menu. The settings existed before, but no notifications were actually sent. See Notifications.