INSPIRATION

AI agents can now change files, update databases, and send requests on their own. But most of them act first and think later, and one wrong action can destroy data or cause damage that cannot be undone. People avoid this by checking before they act. We wanted AI agents to do the same, so we are building PhoenixFlow AI.

WHAT IT DOES

PhoenixFlow AI uses two agents that check each other. The Action Agent plans the task and proposes each step. Before anything runs, the Safety Agent tries the step on a safe copy, finds what could go wrong, lists the side effects, and suggests a safer order of steps. Then the plan runs with a backup, and the result is checked. If something fails, the system goes back to the backup.

Our demo gives the same task to two agents side by side. One agent makes a risky change and breaks things. The other finds the problems first and chooses a safer way.

HOW WE ARE BUILDING IT

We are using Python. The demo uses a fake company folder and a small database, so no real data is ever at risk. The two agents are separate roles powered by an AI model, and a simple workflow passes the plan between them, runs the approved steps, and checks the result.

CHALLENGES WE RAN INTO

Deciding what counts as a risky action, and how to test the actions safely without damaging anything real. We solved this by running every test on a copy of fake data.

ACCOMPLISHMENTS THAT WE'RE PROUD OF

We designed a clear two-agent system where one agent acts and the other questions it, and we planned a side-by-side demo that makes the difference easy to see.

WHAT WE LEARNED

Safety works best when it is part of the plan from the start, not added at the end. Testing on a copy makes risky actions recoverable.

WHAT'S NEXT FOR PHOENIXFLOW AI

We want to protect more tools like email and cloud services, add a human approval step for very risky actions, and give clear reports that explain every decision.