Charlie Tang, Founding Engineer at Fragment, on Building Reliable AI Agents

For AI agents to make real business decisions, they need to do more than perform well in a demonstration. They need to be reliable when faced with the complexity of real-world work. At Fragment, a San Francisco-based applied AI startup, Charlie Tang is leading the effort to test whether its AI agents can be trusted with those decisions. The agents automate procurement work for large corporate customers, where mistakes can have direct financial consequences.

The challenge is not simply to build agents that can complete tasks, but to establish whether they can do so reliably before customers depend on them. Tang's work focuses on evaluating these systems in simulated environments built from real production data, helping the team identify failures before they reach customers. This work sits at the centre of Fragment's effort to make AI useful in business operations.

Tang grew up in the Netherlands and has liked difficult puzzles for as long as he can remember. In high school, he took part in math competitions and did well. He placed fourth in the 2019 Dutch Mathematical Olympiad and won a bronze medal at the 2020 Benelux Mathematical Olympiad.

He then went to Utrecht University, where he finished two degrees, mathematics and physics, in three years. He graduated with honours in mathematics. His academic background gave him a strong foundation for working through complex problems, a skill that would later prove useful in evaluating AI systems.

Trading offered another kind of challenge. The problems were hard, and the pace was very different from academic life. In his second year at university, Tang applied to Optiver, one of the most selective trading firms in the world, and was accepted as a summer intern. He liked the work enough to return full-time at the end of 2023, joining the Amsterdam office as a trader.

He keeps his account of those years short. The firm is demanding, and people there have to move quickly and keep adapting. Tang says it taught him to keep learning, to take ownership of the business, and to make decisions when the outcome is uncertain. Those habits would later inform his approach to building and evaluating AI systems in a fast-moving field.

By early 2026, Tang was looking for a new challenge. Around him, artificial intelligence had changed shape. Only a few years earlier, it meant chatbots. Now it meant coding agents that could write and ship software on their own.

He saw an opportunity to work on a technology that was moving beyond generating answers and towards carrying out tasks. He wanted to contribute to that progress, particularly as AI agents began taking on work with real operational and financial consequences.

He weighed the decision carefully. His job at Optiver was prestigious, and the path ahead was clear. But AI was advancing quickly, and he judged that the opportunity to work at the front of the field was worth pursuing. He resigned without another job waiting.

The decision to leave came first. Finding the right problem to work on came later. While he searched, Tang came across Fragment, an applied AI startup in San Francisco, and the problem it was working on. Procurement is the process companies use to buy the goods and materials they need. It underpins the physical economy, yet much of it is still handled by hand, through repetitive workflows that eat up time.

That seemed to him a natural place for AI agents to help. But automating procurement requires more than an agent that can complete a task under ideal conditions. The system must handle complicated workflows, respond appropriately to unexpected situations, and make decisions that businesses can rely on. In procurement, a mistake can cost real money, so reliability is a requirement rather than an added benefit.

Fragment was already growing quickly, building AI agents for enterprise procurement workflows, and has since raised a $45 million Series A. Tang joined as a founding engineer, taking on work that would help determine whether its agents could operate reliably in real business environments.

Today, Tang works as a founding engineer at Fragment, leading efforts to evaluate the reliability of its AI agents before they are deployed in real business environments. The central question is simple to state and hard to answer: can you know whether an agent is reliable before it acts?

Testing an agent in a controlled demonstration is not enough to establish that it can handle the demands of a real business. Production environments contain complex data, unusual situations, and decisions whose consequences may not be obvious until something goes wrong. An agent that performs well on a narrow set of tasks may still fail when conditions change.

Tang's team addresses this by testing agents in simulated environments built from real production data. These simulations recreate realistic working conditions, allowing the team to examine how agents behave across a range of situations before deploying them to customers. The goal is to uncover weaknesses early and understand where a system may fail, rather than discover those failures after it has begun handling real procurement work.

The use of production data is important because it helps make the simulations representative of the environments in which the agents must operate. The team can test how systems respond to the complexity of actual business workflows, including situations that may not have been covered during training. This makes evaluation more than a final check before release. It becomes a way to measure reliability, identify limitations, and improve the system before customers depend on it.

His view is that how an agent is evaluated matters as much as how capable it is. A powerful system that cannot be trusted will not be given real goods or real money to handle. For businesses considering AI agents, the ability to demonstrate reliability is essential to moving from experimentation to practical use.

AI systems engineering is also a young discipline. In older fields, it takes many years to reach the frontier. Here the frontier is only a few years old, which means the people building production systems today are helping decide how the work is done. Tang says he wants to be one of them. His personal goal is to become a world-class AI engineer and to help set standards for how these systems behave.

For now, his focus is on helping Fragment grow into a leading platform for autonomous procurement. As AI agents take on more consequential business tasks, the ability to test them in realistic conditions will remain important to their adoption. Fragment’s work addresses that challenge by making it possible to evaluate agents against the complexity of real procurement operations before customers rely on them.

Fragment has just come out of stealth with $56 million in total funding, according to reporting from Axios and AI Magazine. The Series A was led by 8VC, while General Catalyst and A* co-led the seed round. Roelof Botha, former head of Sequoia, is also an investor. The funding will support continued development of Fragment's technology, expansion across source-to-pay and deployments with more large enterprises.

Longer term, he sees procurement as a starting point. He pictures the same approach spreading across the physical goods industry, changing how products are sourced, made and moved around the world. Getting there, in his view, comes down to building agents that businesses are willing to trust with real goods and real money.