Watch Parapet stop an agent — the real run.
This is the actual demo: a real Microsoft Agent Framework agent on a local open-source model (Qwen2.5-3B) is told to wipe a customer's records. Play the flow, then flip the switch to see what happens without Parapet.
Everything above reflects a real, unmocked run — real framework, real local model, real tool execution, real Cedar decision. The model makes the identical choice in both modes; Parapet is the only difference.100% real run
The real agent, live — right here.
Not a recording: click a prompt below and watch the same Cedar-governed Microsoft Agent Framework agent decide what it's allowed to do, streaming in real time. Runs on a shared demo instance, so an occasional cold start (~30s) is normal.
Embed not loading? Some browsers block third-party iframes by default.
Open the live demo in a new tab →Every decision flows to the control plane.
The block you just watched doesn't vanish — it becomes something the team can see, review, and turn into a permanent test. These are the actual product screens.
Fleet overview
Every agent, its deny rate, and the most-denied calls across the fleet — execute_shell, export_customer_pii, and the rest — at a glance.
Human review queue
A reviewer flags a decision that looked wrong; it lands here for triage, with the action, the outcome, and everyone's notes — content-free.
Turn it into a test — and catch regressions
One click captures a decision as an eval case. Re-run after any policy edit and Parapet diffs the runs — a deny that a change quietly loosened to allow shows up as a regression.
Per-decision detail
Each case shows expected vs actual and the policy engine's own reason — so a failing control is legible, not a mystery.
