Silverstream AI: The infrastructure for reliable autonomous agents.
Global adoption of autonomous agents requires specialized infrastructure. Silverstream AI exists to build it, focusing on user alignment and accessibility.
“Information sort of wants to be expensive because it is so valuable — the right information in the right place just changes your life. On the other hand, information almost wants to be free because the cost of getting it out is getting lower and lower all the time. So you have these two things fighting against each other.” - Stewart Brand - wiki
Current State of Autonomous Agents
Fully autonomous AI agents, capable of handling any situation without human intervention, are still a distant goal. Much like early cruise control systems in cars, current AI systems are far from reliable, only succeeding about half the time on simple tasks. The development of Level 5 agents, the equivalent of fully autonomous vehicles in the AI world, represents the next significant milestone in this journey.
Achieving robust autonomy demands progress in system engineering, better training and inference time data, agentic models, and diverse evaluation frameworks.
Today, Silverstream AI announced a $1.2M pre-seed financing round led by Google Gradient to create reliable, user-aligned agents.
Our goal is to create reliable and user-first agents. Today, that means providing infrastructure to companies, developers and researchers, as well as sharing data and educational resources on how to use them, all in service of creating useful, user-aligned agents accessible by anyone.
To realize our vision of trillions of agents, we recognize that users must be front and center and in control of their agents. To help bridge today’s biggest gap — high-quality data — Silverstream AI is releasing Pasta-1T, the largest publicly available dataset of agentic interactions, raising the performance bar for baseline models.
The Data Gap and Agents’ Reliability
Working closely with our partners at ServiceNow, NEAR AI, and academic labs, we concluded that reliability remains the chief bottleneck. As large language models rely on extensive, high-quality data, specialized datasets are key to boosting an agent’s success rate. While a few smaller datasets exist, they fall short of training truly autonomous web agents.
A recent ServiceNow AI paper, “Fine-Tuning Web Agents: It Works, But It’s Trickier Than You Think,” shows that performance improves only with larger datasets. However, gathering these domain-specific datasets makes data collection for real-world fine-tuning impractical for individual teams.
Moreover, collecting and curating data at scale is expensive. Many developers and researchers can’t access or afford the breadth of interactions needed to train high-performing agents. We’ve been deploying a network of autonomous agents across the open internet to solve this. Driven by their intrinsic curiosity, these agents navigate the vast complexity of the web, continuously generating agentic interaction data.
The Journey to Agents
Silverstream AI emerged from CEO and co-founder Manuel Del Verme’s research on Hierarchical Reinforcement Learning at Mila. Motivated by a passion for AI as an equalizing force, Manuel spent years advancing RL at Google Brain, Meta Reality Labs, and R&D collaborations across the industry. During his doctoral studies at Mila/McGill under Doina Precup, he studied how agents develop their own tools and coordinate sub-agents, furthering multi-agent systems research. As a longtime supporter of open source, Manuel contributed to efforts like the Farama Foundation to standardize RL interfaces across the community. He previously contributed to the reinforcement learning team of EquiLibre Technologies, which trades a sizable volume of crypto and equities markets. In 2023, he helped publish WorkArena and open-source BrowserGym, a leading web-task automation framework recognized at NVIDIA GTC, ICLR, and ICML.
A New Era of Autonomous Agents
We’re seeing strong interest from scientists for tailored datasets. Developers and researchers are already leveraging our data to launch new projects and push the boundaries of autonomous AI agents.
If you’d like to explore our Pasta-1T dataset and be part of the Silverstream AI’s community, join our waitlist. If you’re curious about potential applications, seeking collaborative opportunities, or want more information, contact us.
