Conference keynotes have pitched the “agent economy” for three years: AI agents that hold money, hire each other, and settle their own bills without a human approving anything. Until recently that idea lived mostly in slide decks. On Aug 29, 2026, the three.ws team published a long engineering post on the Hugging Face blog describing a stack that actually does it, and released everything under an Apache-2.0 license. I opened three.ws this morning to check whether any of it was live. The page loaded in 0.3 seconds, which is already better than most AI product launches I’ve clicked through this year.
What shipped
The team frames an agent as four problems: a body, a brain, a wallet, and distribution. Most agent frameworks quietly skip the first and the last. This one ships all four.
The body is a rigged 3D avatar in glTF format, the kind you can animate in a browser or a game engine. The brain is a multi-model router with persistent, typed memory, so the agent remembers things between sessions instead of resetting every time you close the tab. The wallet is on-chain identity with real keys. Distribution comes in four flavors: a web component you can embed anywhere, an MCP server so other agents can use your agent as a tool, a ChatGPT app, and an Android install.
Nothing here is proprietary. The code sits on GitHub, the models sit on Hugging Face, and a public version endpoint reports the exact commit currently running in production. That last detail sounds small. It isn’t. When a vendor tells you their demo is “live,” you usually have no way to know what build is running or whether it’s the same one from the video. A version endpoint you can hit yourself is the cheapest trust feature I’ve seen all year.
The 3D generation part is free and it works like an ops problem
At three.ws/forge, you type a prompt and get a textured 3D model in GLB format. No account, no API key, no payment.
Under the hood, generation is treated as a routing problem. Every request goes to the cheapest healthy engine, with live health checks and automatic failover: Microsoft TRELLIS handles free text-to-3D through NVIDIA NIM, while Hunyuan3D 2.1/2 and TripoSR run on Hugging Face Spaces for photo-to-3D work. If one engine dies or gets slow, the request moves. Anyone who has run production infrastructure will recognize this as the boring, correct way to build it.
The wallet part is the part people will argue about
Agents in this stack hold their own keys and pay each other in USDC per call, over x402, the HTTP 402 payment standard. No human signs the individual transactions.
I’ll be honest: I don’t know whether anyone wants to pay a 3D avatar by the call. The three.ws post doesn’t pretend to answer that either, which I found refreshing. What interests me is narrower. The plumbing exists, it’s open source, and you can inspect every line of it. For the first time, “agents transacting with each other” is something you can run in two minutes instead of something you hear about at a conference.
There’s a safety layer the team calls a guard chain, described in the post as “autonomy with a seatbelt,” wrapping every autonomous action before it executes. As far as I can tell it hasn’t been independently audited, so treat that as a claim, not a guarantee.
Why this matters if you’re in marketing
The distribution layer is the underrated part for anyone who sells things online.
A web component means an agent can live on any website, not in a chat window. An MCP server means your agent can be a tool that other people’s agents discover and, in this setup, pay to use. The ChatGPT app and Android builds cover the storefront channels. If agents with budgets ever become normal, the question stops being “how do we market to people” and starts including “how do we make our tool discoverable and billable to an agent.” That’s a channel strategy that barely exists yet, and this stack is an early working example of it.
The post is currently trending #1 on the Hugging Face community blog, with around 40 upvotes in a day and 170+ on the team’s related summer report. Attention is not adoption. But in the marketing niche specifically, almost nobody is doing hands-on coverage of agentic payment rails, which is odd, because payments and distribution are supposed to be our department.
Reasons to stay skeptical
Three things keep me from calling this a turning point.
First, upvotes are not production traffic. A working demo and an economy are different things, and nobody has shown agents transacting at volume in the wild.
Second, per-call USDC pricing raises questions nobody has answered. Who sets the price? What stops a loop of agents billing each other into oblivion? The guard chain is the stated answer, but it’s the project’s own answer.
Third, the 3D body might be a gimmick. The wallet, the memory, and the distribution layer would all work fine without an avatar. I suspect the body is what gets the clicks and the wallet is what gets the copycats.
To be clear about my own testing: I verified the site loads and spent time with the post, the GitHub repos, and the version endpoint. I did not run a paid agent-to-agent transaction myself. Everything in this piece is checkable, and the links below let you check it.
What to do now
- Go to three.ws/forge and generate a 3D model from a prompt. It takes about two minutes and requires no account. Judge the output quality yourself.
- Read the Hugging Face post end to end. It’s rare to find an agent writeup where every claim has a repo behind it.
- Hit the version endpoint and compare it to the GitHub commits. Small habit, useful pattern for evaluating any AI vendor.
- If you sell a tool or API, sketch what per-call agent pricing would look like for your product, even as a thought experiment. The x402 standard is worth an hour of your attention.
The agent economy has been a keynote topic for years. As of last week, it’s also a repository you can clone.
Source: Building 3D AI agents end to end, Hugging Face Community Blog, Aug 29, 2026.
