Physix FrontierIs giving AI colleagues a face enough?
I compared commercial products like Coworker with this permissioned coworker framework demo on Hacker News and actually ran it.
Over the past two months, I've used AI Agents to build workflows, with Doubao being one of my main tools; in the last month, I started using OpenAI, and I just picked up Claude a few days ago. Capability isn't the biggest issue; the hard part is whether it can be trusted once it enters real collaboration. By "AI Agent," I mean programs that can break down and execute tasks for humans. A "permissioned coworker" is an AI colleague with defined permission boundaries. If it's unclear what data it can access, whether it can edit documents, or who rolls back changes if something goes wrong, users won't dare to let go. The idea behind this demo is straightforward: first, give the AI colleague a recognizable visual shell.
When you open the page, the first screen shows "The cast," featuring fourteen shapes. Each shape can be inspected and recolored, with soft-body renders alongside. You can understand soft-body rendering as giving characters volume and elasticity. I clicked on a few avatars and changed them to different colors. The effect isn't flashy, but it's definitely useful. For example, an analytical Agent defaults to gray-blue and is read-only only; changing it to a bright color can indicate it's preparing to submit. Pure text status is too easy to overlook; color changes at least allow people to know at a glance if it has crossed the line. I envisioned integrating this into a meeting minutes workflow, where the Agent can only read meeting notes and cannot touch the CRM, with avatar status indicating whether the current action is out of bounds. The direction works, but there needs to be permission switches later, otherwise it's just decoration.
The sticking points are also obvious. I couldn't find configurable permission items. Who can access, who can approve, where task logs are stored—none of this was visible on the public page. It feels more like a design kit than an engineering backend. Making identity visible was prioritized, which is suitable for product prototypes. Fourteen shapes aren't many, but they're enough to distinguish roles, with low visual cost and no reliance on complex 3D. The permission model hasn't been implemented, so it's far from production-ready. If a team has many Agents, fourteen shapes might not be enough, and it could easily become a pretty skin without solving the trust problem. The conclusion is: it depends. It's good for reference by people building AI colleague products, but not suitable for direct adoption in companies. Next, I'll try it with a real workflow, first defining three permission levels: read-only, draft, and submit, then see if avatars and colors can really reduce misoperations.
📌 This article is compiled from Hacker News; original source: https://eisenzopf.github.io/Thelve-coworkers/demo/
All rights reserved by the original authors. This is a compilation and independent analysis based on public reports.
