Agent 2.0.4: see the whole event, not the aftermath
releaseagentperson detectionvehicle detectionpre-rollfalse alarms
Agent 2.0.4 is rolling out now. It’s the largest update we’ve shipped to the intelligence that runs on your own hardware — and if you already run Vigil, you don’t have to do anything: your agent updates itself during its quiet window.
The headline is a new, more capable detection model — YOLOX, a modern on-device detector that recognises people much earlier and more reliably than before, typically within about a second of someone entering frame, including people who are partially visible or still some distance away. It runs on the same on-device runtime as before: nothing new to install, nothing sent to the cloud for analysis.
But better detection isn’t just about recognising a person or a vehicle. It’s about deciding whether something is actually worth alerting you about — and giving you the full context when it is. That’s what most of 2.0.4 is: an event engine around the detector. Four things it now does:
Every motion-clip system has the same embarrassing flaw: by the time motion is detected and recording spins up, the interesting part is over. The clip shows a closing door — not the person who walked through it.
Agent 2.0.4 fixes this properly. Event-only cameras keep a small rolling buffer of the last few seconds on your machine — a few megabytes on disk, negligible CPU, and nothing leaves your network. When an event is confirmed, the alert clip is cut from that buffer: it starts about eight seconds before the trigger. You see someone approaching, not a clip that opens after they’ve already arrived.
The fastest way to ruin a security system is to make it cry wolf. In 2.0.4, an alert requires meaningful activity in the relevant detection region — not simply motion anywhere in the camera view. Wind in the trees next to a parked car no longer produces “vehicle detected”; a shadow sweeping across the driveway is recognised as a shadow and discounted. We tested this on a very windy day, next to a parked car, repeatedly.
Someone who lingers in frame for five minutes is one event — not an alert every sixty seconds. Persistent activity is grouped into a single event, and person and vehicle detections remain independent, so a passing car can never mask a person.
The rolling pre-event buffer lives on your machine, on your network. Only confirmed events are uploaded — so you get the context of pre-roll without turning event-based recording into continuous cloud uploads. If you run Vigil alongside an existing NVR or DVR, this is the part that matters: full evidence clips, event-sized bandwidth.
If you run the Vigil agent today, it updates itself automatically during its quiet window — typically overnight. New installations get 2.0.4 immediately. Open Settings → Agent & machines in your dashboard and you’ll see the version your agent is running.
If you’re not running Vigil yet: your existing cameras, a small agent on a machine you already own, and alerts that show you the whole event. That’s the product. Get started →
Ready to see it? Vigil turns the cameras you already own into cloud CCTV in an afternoon — hosted in AWS Sydney, A$10/camera/month. Start free →