The rise of AI agents has created a new kind of excitement in the software world. People no longer talk only about chatbots that answer questions. They now talk about systems that can browse websites, click buttons, complete forms, gather information, and carry out real tasks across the internet. That shift sounds powerful, but it also brings a hard problem: reliability.
This is where Manuel Del Verme and Silverstream AI enter the conversation. As the co-founder and CEO of Silverstream AI, Manuel Del Verme is working on one of the most important challenges in agentic AI: making autonomous web agents dependable enough for real users, real developers, and real business workflows.
Instead of treating AI agents as flashy demos, Silverstream AI is focused on the deeper infrastructure behind them. The company is building around the idea that reliable autonomous web agents need better data, better evaluation tools, stronger developer infrastructure, and a more open ecosystem.
Who is Manuel Del Verme
Manuel Del Verme is the co-founder and CEO of Silverstream AI, an AI infrastructure startup focused on autonomous web agents. His work sits at the intersection of reinforcement learning, AI agents, browser automation, and developer tools.
His background is important because web agents are not simple software scripts. They need to understand goals, react to changing interfaces, make decisions across multiple steps, and recover when something goes wrong. That kind of work connects closely with reinforcement learning and agent research, areas where Manuel Del Verme has built deep experience.
Before founding Silverstream AI, Manuel’s research path connected with Mila, McGill University, Google Brain, Meta Reality Labs, and other research environments. His work has been tied to topics such as hierarchical reinforcement learning, multi-agent systems, web-task automation, and agent evaluation. This gives him a strong base for building a company that is not only chasing the hype around AI agents, but trying to solve the practical problems that decide whether these agents can actually work in the real world.
That mix of research and product thinking is a major part of the Silverstream AI story. The company is not simply asking what AI agents can do in a controlled demo. It is asking what they need in order to become useful, trustworthy, and scalable.
What Silverstream AI is trying to solve
Silverstream AI is focused on infrastructure for reliable autonomous web agents. In simple terms, these are AI systems that can interact with websites and browser-based tools on behalf of users.
The promise is easy to understand. A reliable agent could help with online research, form filling, internal software workflows, customer support tasks, business operations, and repetitive browser-based work. But the challenge is just as clear. The web is messy.
Websites change. Buttons move. Login flows vary. Forms ask for information in different ways. Pages load slowly. Some tasks require memory, judgment, and multi-step planning. An AI agent may succeed on one website and fail on another that looks only slightly different.
That is why reliability matters so much. If an agent makes one mistake in a long workflow, the whole task can fail. If it clicks the wrong button or misunderstands the user’s intent, it can create risk rather than value. This is especially true for business users who may want agents to work inside complex systems, dashboards, CRMs, internal tools, or enterprise software.
Silverstream AI is trying to address this gap by building the support layer that helps agents become more consistent. The company’s public messaging points toward better data, open infrastructure, evaluation systems, and tools that help developers build agents with more confidence.
Why reliable autonomous web agents matter
AI agents are often presented as the next big step after chatbots. A chatbot can answer a question. A web agent can potentially take action. That difference is huge.
For businesses, autonomous web agents could reduce repetitive manual work. They could help teams collect information, prepare reports, check web-based systems, fill routine forms, or move data across tools. For developers, they could become a new layer for building automation without writing fragile scripts for every website.
But none of this works if the agent is unreliable.
A demo that works once is not enough. A real agent needs to perform across different websites, changing layouts, unexpected pop-ups, incomplete data, and user-specific instructions. It also needs to know when to stop, when to ask for help, and when not to take an action.
This is why Manuel Del Verme is positioning Silverstream AI around reliability rather than just automation. In the long run, the companies that win in agentic AI may not be the ones with the loudest demos. They may be the ones that make agents safer, easier to test, easier to improve, and easier to trust.
How Manuel Del Verme is approaching the reliability problem
The work of Manuel Del Verme with Silverstream AI points to a practical idea: reliable agents need more than a powerful language model.
A strong large language model can understand instructions and generate useful responses. But a web agent must also observe a page, choose actions, remember steps, interact with tools, and learn from feedback. That means reliability depends on a full system around the model.
Several building blocks matter here.
First, agents need better training data. They need examples of real web interactions, including what the user wanted, what the agent saw, what action it took, and whether the task succeeded.
Second, agents need evaluation tools. Developers need to test how agents perform across different tasks, websites, and environments. Without clear evaluation, it is hard to know whether an agent is actually improving or just performing well on a narrow set of examples.
Third, agents need infrastructure. This includes browser environments, orchestration tools, safety controls, debugging workflows, and systems that let developers observe what agents are doing.
Fourth, agents need user alignment. They should serve the user’s goal, respect boundaries, and keep the user in control. A reliable agent is not only one that completes tasks. It is one that completes the right tasks in the right way.
This is the direction Silverstream AI is building toward. The company’s focus on open infrastructure and data shows that it sees agent reliability as a community-wide challenge, not just a private product feature.
The role of data in building better AI agents
Data is one of the biggest barriers in autonomous web agent development. A model cannot become strong at navigating the web if it has not seen enough high-quality examples of web interaction.
That is why Pasta-1T has become an important part of the Silverstream AI story. The company has described Pasta-1T as a large open dataset of web trajectories designed for training web agents. A web trajectory can be understood as a record of an agent moving through a task: what it sees, what it decides, what it clicks, and how the process unfolds.
This kind of data matters because web agents do not learn only from text. They need to learn patterns of action. They need to understand how tasks move from one step to another. They need to recognize when a page has changed, when a task is complete, and when a previous action created an error.
Better data can help close the gap between agents that work in controlled examples and agents that work across the open web. It can also help researchers and developers fine-tune models for specific use cases, such as online research, internal business workflows, or browser-based operations.
For Manuel Del Verme, this data-focused approach connects naturally with a research background in agents and reinforcement learning. If the goal is to build systems that act, then the quality of action data becomes central.
Silverstream AI and the open ecosystem around autonomous agents
One of the notable parts of Silverstream AI’s public positioning is its focus on openness. The company has spoken about a transparent, private, and user-aligned ecosystem for reliable AI agents.
That matters because autonomous web agents are still early. Developers need shared tools. Researchers need strong benchmarks. Startups need infrastructure that lowers the cost of building. Independent builders need access to data and environments that would otherwise be expensive to create.
This is where open datasets and open tooling can help the whole space mature faster.
The broader agent ecosystem already includes important work around web-task automation and agent benchmarks. Projects such as WorkArena and BrowserGym have helped shape how researchers think about testing agents in browser-based and software-like environments. These kinds of tools matter because they move the field away from vague claims and toward measurable performance.
Silverstream AI fits into this larger movement by focusing on the infrastructure layer. The company’s approach suggests that the future of autonomous agents will not be built by model improvements alone. It will also depend on datasets, testing systems, developer environments, safety practices, and shared standards.
Why web agents are difficult to scale
Scaling autonomous web agents is harder than it looks because the web is not a clean, predictable environment. It is full of edge cases.
A human can often handle these edge cases naturally. If a button moves, a person looks around the page. If a form asks for something unexpected, a person pauses and thinks. If a website fails to load, a person refreshes, waits, or tries another path.
For an AI agent, each of these moments can become a failure point.
To scale well, agents need to recover from errors. They need to understand changing page layouts. They need to know when a task is risky. They need permission boundaries. They need ways to verify actions before taking them. They also need clear feedback loops so developers can improve them after failures.
This is why Silverstream AI is not just building around the idea of agents as end-user tools. It is building around the infrastructure that makes agent deployment more realistic. In a business setting, reliability is not optional. A company cannot depend on agents that only work half the time or behave unpredictably when a workflow changes.
Manuel Del Verme’s vision for user-aligned AI agents
The phrase user-aligned AI can sound technical, but the idea is simple. An agent should work for the user, not around the user.
A user-aligned web agent understands the goal, respects limits, and keeps the user in control. It does not take actions beyond its role. It does not assume permission where none was given. It does not hide what it is doing. It should be useful without becoming unpredictable.
This matters even more as agents move from simple browsing tasks into higher-value workflows. If an agent is handling business data, customer information, internal software, or sensitive decisions, trust becomes a core product requirement.
Manuel Del Verme has framed Silverstream AI around this kind of long-term challenge. The company’s focus on transparent, private, and user-first agents shows that it sees reliability as more than technical accuracy. It also includes trust, control, and responsible deployment.
That is a smart position in a market where many people are excited about what agents can do, but still cautious about giving them too much control.
Why investors are paying attention to Silverstream AI
Silverstream AI came out of stealth with a $1.2 million pre-seed funding round led by Gradient Ventures, Google’s AI-focused seed fund, with support from Vento Ventures. That funding gives the company room to expand its team, support open-source developers, and keep building the foundational pieces needed for scalable web agents.
Investor interest in this area is not surprising. Agentic AI has become one of the most active areas in software. Companies are looking for ways to move beyond chat interfaces and create systems that can actually do work.
But the market still has a serious bottleneck. Developers need tools that make agents easier to train, evaluate, deploy, and debug. Enterprises need confidence that agents can perform consistently before they put them inside important workflows. Researchers need better datasets and benchmarks to measure real progress.
Silverstream AI sits directly inside that gap. It is not only building an agent. It is building the infrastructure that could help many agents become better.
For Manuel Del Verme, that gives the company a broader opportunity. If autonomous web agents become a major software category, the infrastructure behind them could become just as important as the agents themselves.
How Silverstream AI fits into the future of agentic AI
The next phase of AI will likely be shaped by systems that can take action, not just generate answers. That does not mean chatbots will disappear. It means AI will become more connected to tools, browsers, workflows, and real-world tasks.
In that future, autonomous web agents could become a common layer across software. They may help users research, compare, organize, fill, move, test, and execute tasks across web-based systems.
But for that to happen, the industry needs dependable infrastructure. It needs better training data. It needs evaluation environments. It needs safety controls. It needs systems that help developers understand why an agent failed and how to improve it.
This is where Silverstream AI is trying to build its place. Its work around Pasta-1T, reliable web agents, and open infrastructure points toward a future where agent development becomes more accessible and more measurable.
That accessibility matters. If only a few large companies can afford to build strong web agents, the technology becomes narrow. If researchers, developers, and smaller teams can access better data and tools, the ecosystem becomes much healthier.
This is part of why Manuel Del Verme’s work stands out. His company is addressing a foundational problem at a time when many others are still focused on surface-level agent experiences.
What Manuel Del Verme’s journey says about modern AI founders
The story of Manuel Del Verme also reflects a broader shift in AI entrepreneurship. Many of today’s most interesting AI companies are being built by founders who come from deep research backgrounds and understand the limitations of current systems.
That matters in a field where hype can move faster than engineering. It is easy to say that AI agents will automate the web. It is much harder to build the datasets, infrastructure, and evaluation tools that make those agents reliable.
Manuel Del Verme is part of a founder generation trying to turn research into practical systems. His path through Mila, McGill University, reinforcement learning, and agent-focused work gives him a strong foundation for building in a technically demanding space.
With Silverstream AI, he is working on a problem that could shape how AI interacts with the web in the years ahead. The company’s success will depend on whether it can help developers and businesses move from impressive demos to agents that work reliably in everyday workflows.
That is the real challenge in autonomous web agents. Not proving that an agent can complete a task once, but building the infrastructure that helps agents complete tasks consistently, safely, and in a way users can trust.