Symbolica AI

We scored 36.08% on ARC-AGI-3 in one day using Agentica. The same SDK driving this score also powers agentica.symbolica.ai, where you can build long-running agents in chat. No sign-up, no code. Try it: agentica.symbolica.ai For Agentica, that's at 36.08% at $1005 total cost... vs 0.25% for $8900 with Opus 4.6 directly. We used the same harness from our previous open source ARC-2 solutions on ARC-AGI-3. Read more: https://lnkd.in/gQE2kfFa Code/Logs: https://lnkd.in/gt8CuqNU

Agentica Beta is live: chat to build long-running AI agents. No code, no login required. Describe a task in plain English, connect your tools, and deploy an agent that keeps working in the background. This demo shows an agent tracking new San Francisco rental listings as they appear. This is just one example, you can build anything. Links below

Today we set a new ARC-AGI-2 public eval SOTA: 85.28%. We did it with a simple Agentica agent (~350 lines) that writes and runs code. Other agent harnesses keep everything in the model’s context window. We don’t. Agentica uses a stateful REPL to manage context. This is an RLM-style loop: swap the model, keep the loop. Agentica is actually recursive, unlike traditional RLMs (which are limited to depth=1). Agentica spawns sub-agents which spawn sub-agents, to explore multiple hypotheses and keep context clean. These recursive sub-agents are what got us to SotA. The REPL makes correctness easy to enforce: run, sanity-check, debug, rerun. In our SotA ARC run, sub-agents verify results in the REPL before submitting, instead of shipping a guess. Links in comments below

Today we’re launching Agentica: an open-source SDK (Python + TypeScript), plus the Symbolica Platform for running typed, multi-agent workloads in production. Out-of-the box, our framework goes beyond code mode to re-think how agents interact with their environments. Doing this led to a first-try 77% SOTA score on BrowseComp, with only a few hours of an engineer's time building. We’re also previewing our longer-term work on a neuro-symbolic foundation model architecture using category theory to bridge symbolic structure with neural adaptability. Existing stacks are brittle: tool calls fail, workflows lack types and structure, state gets messy, and reliability collapses as tasks get longer. Agentica addresses this by making agents typed programs inside your codebase, with orchestration, state, security, and observability built for production. With Agentica, teams stop hand-building orchestration diagrams and glue code. You pass your existing functions, APIs, and SDK methods and agents call them directly in execution. Agents can write and execute code, discover tools, spawn specialised sub-agents, coordinate and share state, hand off work, and adapt execution paths at runtime. Run Agentica yourself or deploy on the Symbolica Platform to run dozens of concurrent agents with governance-grade tracing and debugging. When workflows are well-typed code, systems can be generated, edited, and composed far more reliably than brittle GUI flows. Read more: https://lnkd.in/ecy2dCeT -- Khosla Ventures General Catalyst Abstract Day One Ventures Buckley Ventures -- (for BrowseComp benchmarking: all agents had access to a document get tool and a retriever, where Opus 4.5 used a Qwen3-Embedding-8B, while GPT-5 standalone and Agentica + GPT-5 used a Mixedbread)

See you at V2!

Last week, we brought together our SF team, investors, and research community to share a preview of Symbolica's newest product, launching soon. Stay tuned. -- Khosla Ventures General Catalyst Abstract Buckley Ventures Day One Ventures Vinod Khosla Christopher Kauffman Alex Davidov Masha Bucher Kanu Gulati Eric Johnson Vijay Karunamurthy Kevin K. Rooney Sam D'Amico

If you're a Rust developer interested in formal logic, automated theorem proving, types, compilers, and AI please join us for a hackathon at our San Francisco office on Sat, Nov 8th! RSVP now before spots fill up 👉 https://luma.com/1xa9d6nr

Stephen Wolfram stopped by our London office last week to dive into symbolic AI, type theory, formal logic, and what’s next for building reasoning machines. We're hiring across London and San Francisco – apply here 👉https://lnkd.in/eCUhDugH #LifeAtSymbolica #ArtificialIntelligence #AI #AIResearch #SymbolicReasoning #CategoryTheory

Wrapped up a fantastic day at the Oxford University Career Fair! 🎓 Loved chatting with so many brilliant grads about the blend of pure math and AI – and how they can help shape that journey with us. Taliesin Beynon (our Research Director), Victoria Klein (ML Researcher), and Yaroslav (Yarik) Doronin (pretending to be an ML Researcher) will also be at the Cambridge Graduate Jobs Fair on 24 June – come say hi 👋 #LifeAtSymbolica #ArtificialIntelligence #AI #AIResearch #SymbolicReasoning #CategoryTheory