AgentGlass

Imagine spending hours designing and 3D-printing a custom holder for a metal part. Everything fits. Then the factory worker tries to tighten a bolt. The wrench can't get in. That tiny mistake means another design, another print, another delay. That's why I'm building FixtureFit Gemini 4 Argon hackathon with @GoogleDeepMind in San Francisco. The idea: show AI the operator's task, generate the fixture, check the tool's movement, and automatically repair collisions. In my example: Draft 1 by Argon clears the Ø18 socket: +3.98 mm A bigger Ø28 socket: blocked by 1.03 mm After one Argon repair: +1.97 mm (needs 1.5 mm) The important part? Argon generates the design. My independent checker decides whether it passes. The engineer gives final approval. Find the blocked wrench before the print. Built with Gemini 4 Argon and Gemini Managed Agents. @googleaidevs @GoogleAIStudio @AgentGlassInc @DemisHassabis @GoogleDeepMind @GeminiApp @GoogleAIStudio @ivanleomk @OfficialLoganK @joshwoodward @DynamicWebPaige @cerebral_valley @Google nitter.cf/AgentGlassInc/status/2…

I built Clearbell, and it won a prize at a Halloween hackathon in San Francisco. It is for the night-shift nurse in charge at a care home. In a simulated care home, one note says the Room 9 floor is dry. A later note says the tiles by the sink are still wet, and the resident uses a walker at night. Clearbell keeps both notes word for word and lists the checks nobody has done yet, each with its source and who does it. For adults 65+, falls are the leading cause of injury death (CDC). About 1.24 million residents live in 14,742 CMS-certified nursing facilities (KFF analysis of CMS data). Thank you @AWSstartups for the Builder Loft, and @parasail_io, @crewAIInc (Lorenze Jay Hernandez @lorenzejay) and @onedotnew for the talks and the tools Clearbell runs on. The video: Clearbell on the big screen at the event, and the win. Demo: youtu.be/1KGgGaXYleM Story film: youtu.be/p1iqVCcKUhM Partner talks: youtube.com/playlist?list=PL… @AgentGlassInc nitter.cf/SidraMiconi/status/211…

Care-home tools count alarms raised and tasks logged. Does the NEXT CHECK survive a contradiction, and a real outside fact? Clearbell is for the night-shift nurse in charge. In Maple Hollow, a simulated care home, every room, note and person is invented. At 21:40, an aide: “Mopped bathroom floor in 9. Sign up. Dry.” At 22:05, Elsie: “Sign in 9 was on the floor. Tiles by the sink still wet.” The resident uses a walker at night. Clearbell keeps both notes word for word and lists the checks nobody has done yet, each with its source and who does it. Outside evidence can add work, never remove it. The board answers one question: what still needs checking, and who does it. Impact: falls are the leading cause of injury death for adults 65+ (CDC); about 1.24 million residents in 14,742 CMS-certified nursing facilities (KFF analysis of CMS data). Innovation: an outside fact changes the work. You.com (@youdotcom) finds a real FDA Class 1 recall for the bed on file, the work order changes, and a person approves it before @onedotnew files it. Technical: a @crewAIInc crew and Flow with a human pause and a night-shift router; @parasail_io text and vision models; You.com Search → Contents; One files a real Linear issue; hosted on @awscloud (EC2 behind CloudFront). Experience: the Room 9 board, try-your-own-notes, onboarding, a 6-slide deck in the product, four videos. Docs: a README with visuals and all four videos; one command runs it with no keys. Live (AWS EC2 ,CloudFront, etc): d4wtf1to1i5mh.cloudfront.net… Story film: youtu.be/p1iqVCcKUhM Built by @SidraMiconi @AWSstartups nitter.cf/SidraMiconi/status/211…

"Dry." Then: "Tiles by the sink still wet." Clearbell keeps both notes, and lists the checks still undone. Why it matters: I built Clearbell for the night-shift nurse in charge at a care home. The resident in Room 9 uses a walker at night. Keeping an unresolved wet-floor report visible helps the nurse see what still needs checking. For adults 65+, falls are the leading cause of injury death (CDC). Across 14,742 CMS-certified nursing facilities, about 1.24 million residents receive care (KFF analysis of CMS data). How I built it: I used @parasail_io models to read the notes and a vision model to read a corridor frame as "possible water". No people appear in the frame, and frames aren't kept. I used You.com Search to find, and Contents to read, a real FDA Class 1 recall for the bed on file. That recall changes the work order. I made a @crewAIInc crew and Flow that pause until Noel, the duty administrator, approves. I used @onedotnew to file the approved request as a real Linear issue, with a second press blocked. I hosted it on AWS, on EC2 behind CloudFront. One command runs it with no keys. I built it at the @AWSstartups Builder Loft in San Francisco for a Halloween hackathon. Maple Hollow is simulated: every room, note and person is invented. 0:00 Two notes disagree 0:30 Checks with their sources 0:58 A real FDA recall changes the work 1:23 Noel approves, One files it 1:34 The night, rules first 1:45 Camera report 1:55 Tested blind Full demo: youtu.be/1KGgGaXYleM sidramiconi.com AgentGlsss.dev @AgentGlassInc

A 3D-printed fixture can hold the part perfectly and still block the operator's wrench, and you usually find out at the machine, after the print. FixtureFit checks that the wrench clears the fixture before anything is printed. Here the socket grows from Ø18 to Ø28 mm and the wrench hits the holder: −1.03 mm, FAIL. Two forks of one saved Gemini Managed Agent try fixes in parallel, and the app's own checker decides the verdict, not the model. One repair: +1.97 mm, PASS. Here the relief fork passed on its first try, and 3 of 3 measured live runs passed. Then it waits. Nothing moves until the engineer signs off. I built it with Gemini 4 Argon and Gemini Managed Agents, at the hackathon run by Cerebral Valley and Google DeepMind. @AgentGlassInc @DemisHassabis @GoogleDeepMind @GeminiApp @GoogleAIStudio @googleaidevs @ivanleomk @OfficialLoganK @joshwoodward @Google @DynamicWebPaige @cerebral_valley nitter.cf/SidraMiconi/status/211…