Sandboxed code steps and the Agent Harness

MagOneAI v3.0 is a major release. Workflows can run your own code safely, agents can run on their own without a workflow, and a new Agent Harness lets you see everything an agent did. Knowledge bases, model controls and file handling are all improved, and we fixed several important security issues.

Highlights

Code node: run Python as a workflow step

A workflow is a series of steps, called nodes, on a visual canvas. The new Code node runs your own Python code as one of those steps. Use it for calculations, reshaping data between steps, or any logic that's easier to write as code. Each run happens in its own isolated, single-use sandbox that can't see other runs or the platform itself.

Why it matters: When you need custom logic, you can write a few lines of Python instead of chaining together several other nodes.

Learn more: Code node

Standalone agents and the Agent Library

Most agents run as a step inside a workflow. You can now also run an agent on its own, for example to test it, answer a one-off request or process a single document, and its results are saved like any other run. The Agent Library is where you keep reusable agents.

Learn more: Standalone agents and Agent Library

The Agent Harness

The Agent Harness records what agents do during a run:

• An Artifacts view shows what an agent produced and used along the way, such as search results and fetched pages.
• An action ledger tells an agent what it has already done in this run, so it doesn't repeat work.
• The Blackboard is a store for each run where agents keep intermediate results that later steps can look up.

Why it matters: When an answer is wrong, you can see exactly how the agent got there.

Learn more: Execution artifacts and the Debugger

Improvements

• Knowledge base overhaul. A knowledge base is a collection of your documents that agents search to answer questions. You can now keep a knowledge base in sync with a SharePoint folder, and add content from a web address or by crawling a website. Extraction of text from PDF and Office files has been rebuilt to handle each page in the most suitable way. See Knowledge bases and RAG.
• More control over AI models. Set token limits per person and per model. Choose which model each agent, or each individual step, uses, and pick a different model for a single chat message. See Model configuration and Usage & quotas.
• Files in chat. Upload files to a chat and see them displayed, including files an agent creates while it works.
• Your brand, your platform. Replace MagOneAI's name and visual identity with your own across MagOneAI Studio, MagOneAI Hub and the login pages. See Admin Portal.
• Project-specific agents and skills, so each project can keep its own.
• Undo and redo on the canvas.
• Multiple database connections per workspace. See Database tools.

Security

• A critical finding from an external security test was closed. A monitoring service had been publicly reachable. It was rated 9.8 out of 10 for severity and is now locked down.
• Data isolation tightened. Gaps that could have let one user or organization access another's files or search results were fixed.
• Protection against malicious Office files. Uploaded Office documents that are designed to expand to an enormous size when opened are now rejected. See Input limits.