Opus 5.5 message router: 10,000 decisions visualized as glowing strands

Amir @Amir_RadaritOriginally posted 2026.10.30

Open original on X

This video looks like art, but it's actually the most practical thing I've built with Opus 5.5 so far. Every glowing strand is a decision. Here's the problem it solves. Every day I get a flood of small things: comments, DMs, emails, ideas, tickets, random links people send me. Each one needs a tiny decision. Reply? Research it? Ignore it? Is it a scam? Does it need me personally? Individually they're nothing. Together they ate hours of my day. So I stopped answering each one and built a router with Opus 5.5 instead. It works in layers, top to bottom, just like in the video: Sources: everything that comes in, comments, mail, DMs, tickets State: it figures out the context, who sent it, what it's about, what happened before System One: three typed questions on every single item. Pick a choice, give it a score from low to critical, and catch anything unusual Routes: then it sends the item down one of six paths: research, write, review, escalate to human, quarantine, or just leave it alone Act: the matching agent does the work, and Opus 5.5 is the writer at the end of the write lane 10,000 decisions in this run. p50 around 250 ms per decision. Total cost: $0.42. That last number still makes me laugh. Ten thousand micro-decisions for less than a coffee. The panels on the right are the ones I actually care about. Spend guard with a hard limit. And an outcome check that has to pass every time: a draft exists nothing gets posted automatically spend stays under the limit the human gate is armed the outcome gets verified "No auto-post" is my favorite rule. The system can draft a reply, but it can never publish it on its own. Anything sensitive goes to escalate_human, which in this run was about a third of the traffic. Phishing-looking stuff goes straight to quarantine. Boring stuff stays flat and doesn't waste a single token. My honest opinion: Most people use AI like a smarter search box. You ask, it answers, you ask again. That doesn't scale, because you're still the one making all 10,000 decisions. The real shift is letting the model make the small decisions and only bringing you the ones that actually matter. And Opus 5.5 is good at this specifically because it's consistent. Item 9,000 gets the same three questions and the same routing rules as item 1. It doesn't get tired at 6pm, it doesn't skip the quarantine check because it's in a hurry. But I want to be clear: the human gate isn't decoration. I don't want a system that acts on my behalf without me. I want one that sorts my day so I spend my attention on the 30% that needs a person, not on the 70% that doesn't. Comment "ROUTER" and I'll DM you the whole setup: the three System One questions, the six routes, and the outcome checks I use And tell me in the replies: if you had a router like this, what would you send straight to quarantine first? For me it's every "quick question" DM with a link in it

This post quotes another post: x.com/takerovitch/status/2113299974586773981