Agentica

We set a new ARC-AGI-2 SotA: 85.28% using an Agentica agent (~350 lines) that writes and runs code.

ARC Weekly Meeting — Sunday, March 1 at 1 PM EST Excited to dig into it and see how it works! Special Guest: Symbolica Agentica Team Join: discord.gg/MSQay8r7mg

@fchollet @fchollet x.com/agenticasdk/st…

@agenticasdk 1. Is it seeing each game only once? (it is of course possible to brute-force any game given infinite trials, but that is not the goal here) 2. Is it using a number of actions per game comparable to what humans need? (upon seeing the game for the first time)

We have now solved all publicly available ARC-AGI-3 puzzles.🧩

We've released the code used to solve all 3 publicly available ARC-AGI-3 games. github.com/symbolica-ai/A…

Join us this Thursday, 26th in London. We’ll be speaking about ARC-AGI-2 and ARC-AGI-3 at the London AI nerd meetup: what’s new, what’s changing, and what it means for agents. @its_hapenin @FireworksAI_HQ luma.com/z4i401h6

@fchollet @vkhosla @fchollet x.com/agenticasdk/st…

@vkhosla @agenticasdk You should try ARC-AGI-3 (developer preview is available now, full benchmark coming in a few weeks)

Well well… ARC-AGI-2 (François Chollet’s “hardest” benchmark) is starting to smell like toast. 🍞🔥 @agenticasdk just set a new SOTA: 85.28% with an Agentica agent (~350 lines) that writes & runs code. Best part: it’s not ARC-specialized—it's a general system that’s strong across other benchmarks too. Details at symbolica.ai/blog/arcgentica What benchmark should we throw at it next?

@GregKamradt Thanks @GregKamradt, looking forward to the rest of the games! x.com/agenticasdk/st…

Really cool to see teams already working on ARC-AGI-3 We'll have public replay links for every online run

@scaling01 x.com/agenticasdk/st…

Watch the replay here three.arcprize.org/replay/ft09-9a…

ARC-AGI-3 ft09 solved in 346 steps

I tried the ARC-AGI-3 test set, it’s quite well done, but I am sure AI models will solve it before year's end. It looks like it began before it was released ☺️. I suggest they already plan to release ARC-AGI-4 soon!

@scaling01 x.com/agenticasdk/st…

Earlier today I wanted to doom about Gemini 3.1 Pro completely failing ARC-AGI-3. Turns out this was due to a bug in the config introduced by GPT-5.3. It was still calling Gemini 3.0 Pro instead of 3.1. I fixed it, made the harness simpler and spend $120. Performance of Gemini 3.1 Pro is much better than the almost random performance of 3.0 Pro. Gemini 3.1 Pro can actually solve some games.

@scaling01 x.com/agenticasdk/st…

Many people have asked us: what changes when an agent has access to a persistent Python runtime? We ran a side-by-side comparison to demonstrate: Agentica's Python REPL-based agent vs traditional tool calling agents Full breakdown below 👇

👀

Vic (@its_hapenin) will chat ARC-AGI-2 results at Engineering Night London - what worked, what didn’t, and what surprised us. With Agentica: • 85.28% with Opus 4.6 (120k) High • +10pp GPT-5.2 (XHigh) vs CoT • +20pp Opus 4.5 vs CoT See you there. Register: luma.com/w5keu427

Funnily enough I tried to dabble with ARC AGI before and with very little success… Super cool to well designed RLMs achieving SOTA :)

Reasoning. Persistent state. Recursion. In our ARC-AGI implementation, agents autonomously decide when and what state to pass into a sub-agent’s REPL, allowing them to focus on analysing training examples, test inputs or both. Check out the logs github.com/symbolica-ai/a….