AI

Sail Research raises $80M to optimize long-horizon AI agents

Inference startup Sail Research Inc. today announced that it has raised $80 million in funding at a $450 million valuation.

The company received the bulk of the capital in the form of a Series A round led by Sequoia. It earlier raised a seed round led by Kleiner Perkins. Sail Research also counts Intel Corp. Chief Executive Officer Lip-Bu Tan, Alphabet Inc. chair John Hennessy and Redpoint Ventures among its investors.

Sail Research operates a cloud platform that developers can use to run long-horizon artificial intelligence agents. According to the company, its infrastructure enables agents to tackle tasks that take upwards of weeks to complete. Furthermore, Sail Research claims that it can run such workloads at a fraction of the price charged by competitors.

The company says that its platform is powered by customized versions of several open-source inference engines. An inference engine is a tool that lowers the hardware usage of AI models. One of the most widely used tools in the category is vLLM, which partly owes its popularity to an algorithm called PagedAttention. The algorithm speeds up inference by enabling AI models to make more efficient use of graphics’ cards built-in memory.

Sail Research runs AI agents in Linux-based virtual machines called Sailboxes. Developers can customize each virtual machine by installing an image, a bundle of software modules and configuration settings. Furthermore, the platform makes it possible to link multiple Sailboxes into an AI agent ensemble. 

An AI agent completes long-horizon tasks by breaking them down into small steps and completing those steps one after one another. Some steps require the agent to wait for an external system to fetch data. According to Sail Research, its platform makes it possible to shut down AI agents while they’re waiting and thereby lower infrastructure costs.

The company evaluated its platform using a benchmark called BrowseComp-Plus ahead of today’s funding announcement. The test measures AI agents’ ability to perform complex online research tasks that take a significant amount of time. According to Sail Research, its platform set a new high score of 90.72% while incurring one tenth the inference costs of rival services. 

“Unlike a human waiting at a keyboard (top priority: speed), agents need scale, reliability, and sustainable cost,” Sail Research co-founder and Chief Executive Officer Neil Movva wrote in a blog post. “Sail finds this efficiency everywhere in the stack: we carefully choose our chips, write custom inference engines, and run a global controller that fully utilizes every computer in our fleet.”

Sail Research will use its newly raised funding to enhance its inference infrastructure. 

Photo: Sail Research