Meet your teammate, Simular 1.0

Today, we’re launching Simular 1.0, our native desktop agent that can perform complex, multi-step tasks across the applications you already use. And for me, this launch is the culmination of a journey that began long before “AI agents” became a buzzword.

In 2019, when I was researching multi-agent systems at DeepMind, few were excited about the path I was exploring except for one colleague, Jiachen, who later became my co-founder at Simular. Fast forward to 2025: the term "AI agent" appears everywhere, and every startup claims to be building one.

But there’s an uncomfortable reality: most “agents” today don’t actually behave like agents.

They fail at executing long, complex workflows. They hallucinate and can’t repeat their own success. And many of the most popular “agentic” use cases are really just generative AI – making decks, videos, or images – tasks that hardly require long-horizon planning and adapting to ever-changing real-world websites and desktop environments.

For us, real agents should understand user intent, initiate tasks, find the best path forward and reliably repeat its previous successes, all on its own. Like a good teammate. The conventional path for building agents is to remove humans. We believe humans are necessary to imbue machines with the right goals and values, make judgment calls, and be gatekeepers of quality. 

And our AI agent isn’t purely built on LLMs. We use a neurosymbolic framework that combines the creative, exploratory nature of LLMs and the deterministic aspect of code, which repeats success and guarantees reliability.

A desktop agent that supports and learns from you

Simular 1.0 is designed to bring AI agents into real, everyday workflows. It’s trained with humans in the loop, allowing users to redirect or correct the agent at any step using natural language. Over time, it learns from this supervision just like a real teammate would, becoming increasingly reliable at repeating successful workflows.

Simular 1.0 also supports contextual task triggers that users can configure so the right workflow launches automatically based on real-time activity on their computer. Bigfoot, our fan-favourite mascot for Simular, surfaces friendly tips and alerts for these task triggers, giving users a more approachable and human feel.

Over the past year, we’ve taken technology straight out of our research lab and turned them into products used by businesses, ranging from mission-critical insurance workflows to open-ended web navigation tasks in creative marketing. Now we’re opening this capability to the broader consumer base.

Is the agentic industry overcrowded? Yes. But the wave isn’t retreating. If anything, it’s accelerating. Just eight months ago, our agent achieved a 34.5% success rate on OSWorld, the benchmark for computer-task execution. Today, it’s at 69.9%, nearing human performance at 72%. As agents move to reliably performing real computer tasks, we expect an explosion of new workflows and use cases that simply weren’t possible before.

Underneath the research and benchmarks is our goal.

Computers were created to help us, not to consume our days with clicking, typing, and tab-switching. Yet that’s what modern work has become. At Simular, our mission is to give people their time back. Time for real passions and their loved ones.

Simular 1.0 is a step toward that future. Thank you for being here with us at the beginning.


Ang
December 2025