
This is an early beta. I'm still actively developing it, and I'd like to work with you on as many AI apps as I can, integrating and improving this together as we go. It's Apache 2.0. Repo: github.com/approximatelabs/a… Site: affordance.rs (13/13)

Approximate Labs

This is an early beta. I'm still actively developing it, and I'd like to work with you on as many AI apps as I can, integrating and improving this together as we go. It's Apache 2.0. Repo: github.com/approximatelabs/a… Site: affordance.rs (13/13)

And when you do want chat, it's exact. I rebuilt the chat formats of 740 models as Affordance apps and checked each one against its official Jinja template and tokenizer. The tokens match. (12/13)

Chat is only one way to lay out a model's context. An Affordance app can lay it out however the task needs: a document, a dashboard, a page the agent rewrites itself. It doesn't need a chat template or a tool template. (11/13)

And something new: an agent can rewrite its own harness while it runs. In Affordance the harness is an app, and an app's source can be state. So what the agent reads, and the tools it has, are things it can change. (10/13)

Affordance also powers multiplet.ai, a fleet framework I'm releasing soon. It runs over 2,000 agents on a single machine, with the VMs, the harness and inference all on that machine. (9/13)

OptMem is @VictorTaelin's tree-based memory, a way of folding context. I implemented it in 120 lines of TSX, and here's how well it works: affordance.rs/showcase/memor… (8/13)

You can also build the more unusual kinds of harness. Recursive Language Models, from @a1zhang, took 116 lines of TSX. It's here: affordance.rs/showcase/rlm (7/13)

I also used it for RL training, which needs the exact tokens the model wrote. I found the usual Jinja round trip doesn't give those back, often dropping tokens in tool calls. Affordance targets tokens directly, so there is no round trip. (6/13)

And here it is inside a coding agent. pi has about 8,200 lines of TypeScript for its agent session. I replaced that with a 312-line Affordance app. It survives a restart, and you can try a new memory technique by editing the app. vs. Pi Durable: affordance.rs/showcase/pi-du… (5/13)

Here's Affordance inside a model server. I put it into SGLang in place of the template and tokenizer step. On Qwen3.6-35B-A3B with prompts over 120k tokens, the first token came about 24% sooner: 421 to 422 ms against 552 to 562 ms for stock SGLang. (4/13)

An app is a TSX file, a lot like a web page. Text becomes tokens, state goes in a box, and a tool is just a function. Layout goes in a style sheet with budgets in tokens. Here one rule caps a test log at 92 tokens, and one more line keeps the failing test at the end. (3/13)

A browser renders a page into pixels, you click, the page changes. Models don't get that today. Their prompt is built as text, then cut into tokens, and the app never sees them. Affordance renders the page straight to tokens, and the model acts by calling a tool on it. (2/13)

I rebuilt the agent harness the way a browser runs web pages, and today I'm open-sourcing it! Affordance is a browser for models: you write an agent as an app, and it runs the loop and renders the exact tokens the model reads. (1/13)

Send me a DM if you want to try it out!

Keep finding more evidence that communicating fleets actually work! This is getting more and more exciting everyday as I push and test the ideas

Swarms are amazing. I just ran a multiplet fleet to harden multiplet itself: literally hundreds of findings (many duplicates), including some legitimate high severity things I've missed in moving fast this past week. Standard agent workflows take too long to find all of these.