Project Perception in practice
Aug 5th, 2026 by Jesper Kråkhede
On the 3rd of August Project Perception entered preview. It is a limited public preview, so only a selected group of customers have access for now. Here is how it actually works.
It lives in the Defender portal, with sections for an overview, chat, sessions, agents and playbooks.
Playbooks are reusable templates that describe how agents should handle a specific scenario. You start a playbook, or you describe the task in chat and Perception finds the right playbook, fills in the inputs and runs it.
A session is a piece of agentic work. You can watch work that is running or recently finished, see what started it, approve or reject actions when the agents ask, stop a session that is heading in the wrong direction and give the agents new guidance.
That supervision is the important part. It turns ‘the AI did something’ into ‘the AI proposed something, and I decided’. Every session is a record of what was done and why.
My advice if you get access: start with the red agents on a limited scope, so you see which attack paths they find in your environment. Then let the blue agents investigate incidents in parallel with your analysts and compare. Only then let the green agents act, and start with approval required for everything.
Measure as you go. How often are the agents right? How much time do they save? Where do they need more context?
Treat it like a new team member. Watch it closely before you trust it.