WWDC26 is great, but watching the videos of How Apple decided to squeeze 20B model and I stumble on their Architecture, Looks very similar to what we shared with them. Left ours @autohandai Mneme architecture, right Apple AFM design Yes, we have been built our accelerators for Small Dense models to run in some of our customers that has Apple servers.

git worktree weren't designed for 10x agents on one repo or mono repos. Each one is a full checkout plus a branch, and the merge only sees text. I've been building cpsg (Causal Program State Graph): agents record changes into a graph stored in git. @autohandai mission is to build infra layer of the modern age where 96% of the code is created, maintained and updated daily by coding agents. For that we need to reinvent the wheel. I'm happy to share early results of our benchmark after months on development. Still alpha but I'm conscious that is still early. I had to first build our agent in RL, then our models, then this. Workspaces are disposable copy-on-write clones. Creating a workspace, 10k files: git worktree: 1,189 ms cpsg: 242 ms 10 at once, 30k files: 22.6 s → 3.8 s It also catches what git merges silently: agent A changes a function's signature, agent B calls the old one from another file. cpsg stops B and names the symbol. I'm inviting a selected group of innovators to try this with their favourite harness @badlogicgames has 1st spot.

Heads up: We’re enforcing mandatory 2FA for all free and paid accounts starting next week. Why? To block automated dots, muses, botnets, eliminate security debt, and keep our infrastructure clean and fast for legitimate developers. Set up your TOTP or Passkey in your settings now to avoid any interruption.

I still think that autohand --deep-research "durable objects from cloudflare " --handoff codex|claude or vice versa is the best way to find things and implement. Works unbias with Fantail. I just learnt about @badlogicgames Pi is now part of @Cloudflare durable objects, I hope they sponsor you at least a nice coffee. Everyone now is a cloud clanker. Results are in the here gist.github.com/igorcosta/8f…

Decision models in llama.cpp are now available The `/v1/systemone` endpoint is available in the latest llama builds. Use it to do Jev-style inference locally, efficiently and privately. Multiple open models are supported with more to come. huggingface.co/blog/ggml-org…

hey @vkhosla The struggle is real. I can confirm. At @autohandai we easily beated @FactoryAI in multiple fronts with our github.com/autohandai/code-c… We're very popular #6 against #9 on Zed Editor and many other editors with $0 raised.

I encourage anyone to read annual report from NVIDIA. @elonmusk basically outbid compute from anyone for the next 24 months. And @AnthropicAI is relying on that. The chances of successful IPO for @AnthropicAI is close to zero.

Introducing Autohand Profiles, claim yours now developers.autohand.ai/@you

My Autohand Code profile developers.autohand.ai/@igor…

You can only fill what is empty.

New beginnings

2027 is the year of small, highly specialised SOTA models running at the edge.

Data you can actually get along with it.

The small team from New Zealand that keep shipping.

A new benchmark index is born, Tokenspeedmaxxing.

Everything I did to orchestrate coding agents was just to proof consensus in organized learning and trajectories that mimic the humans in an office environment. Oh boy, I was wrong half of the time. Then I moved into biology and self reflection and then into mutation engineering arxiv.org/abs/2602.07072, nature is a master in this game, I’m just a copy cat of Nature incredible ability to transfer information from 🧬 strands into a new life form in the most basic effort way. Right now, I’m once again wrong! Works well but if I truly believe that I will unlock the power of our creation into adaptive continuous engineering (ACE), I have to add disorder to drive changes, and once again nature did a chess check on me, quantamagazine.org/disorder-… once more. So, this week I hired the biggest linodes servers money could buy and connected them together on a 10Gigabit network with one single goal to replica disorder on coding and distributed systems,Autohand code and 19 other coding agents to tackle one of the most difficult problems in our category that I’ve been dealing over the past months, migrate all Linux kernel from original code into rust programming. No consensus, running our local models, using trajectory’s memory as tensors for our LoRA adaptors. There are a few rules for the to follow: - they have to be flexible enough to not block out each other. Which me luck. I will write my findings, once the experience finishes. I have been chatting with a few people in the community about this,

You should apply if you’re in the Bay Area. Sid is a good human being.

22 Sept 2026 - LLMs now have the level of expertise as a level 200 for general coding tasks. Still 70% that's the acceptance today.

Despite all the wars going on; This chart brings me hope in humanity and our ability to secure the future. ourworldindata.org/grapher/c…

I spoke to my wife this weekend about this.

We're making easier for Developers and Constructors to build using on our platform. A few updates from us: thread. Our Playground is now GA across all users, free, Pro and Max. autohand.ai/signup