Charlie Tang, Founding Engineer at Fragment, on Constructing Dependable AI Brokers

For AI brokers to make actual enterprise selections, they should do greater than carry out properly in an illustration. They must be dependable when confronted with the complexity of real-world work. At Fragment, a San Francisco-based utilized AI startup, Charlie Tang is main the trouble to check whether or not its AI brokers might be trusted with these selections. The brokers automate procurement work for giant company prospects, the place errors can have direct monetary penalties.

The problem is just not merely to construct brokers that may full duties, however to determine whether or not they can accomplish that reliably earlier than prospects rely on them. Tang’s work focuses on evaluating these techniques in simulated environments constructed from actual manufacturing knowledge, serving to the workforce determine failures earlier than they attain prospects. This work sits on the centre of Fragment’s effort to make AI helpful in enterprise operations.

Tang grew up within the Netherlands and has favored tough puzzles for so long as he can keep in mind. In highschool, he took half in math competitions and did properly. He positioned fourth within the 2019 Dutch Mathematical Olympiad and gained a bronze medal on the 2020 Benelux Mathematical Olympiad.

He then went to Utrecht College, the place he completed two levels, arithmetic and physics, in three years. He graduated with honours in arithmetic. His tutorial background gave him a powerful basis for working by complicated issues, a ability that may later show helpful in evaluating AI techniques.

Buying and selling supplied one other form of problem. The issues have been onerous, and the tempo was very completely different from tutorial life. In his second yr at college, Tang utilized to Optiver, probably the most selective buying and selling companies on the earth, and was accepted as a summer season intern. He favored the work sufficient to return full-time on the finish of 2023, becoming a member of the Amsterdam workplace as a dealer.

He retains his account of these years quick. The agency is demanding, and other people there have to maneuver shortly and hold adapting. Tang says it taught him to continue to learn, to take possession of the enterprise, and to make selections when the end result is unsure. These habits would later inform his strategy to constructing and evaluating AI techniques in a fast-moving subject.

By early 2026, Tang was searching for a brand new problem. Round him, synthetic intelligence had modified form. Just a few years earlier, it meant chatbots. Now it meant coding brokers that would write and ship software program on their very own.

He noticed a chance to work on a know-how that was shifting past producing solutions and in the direction of finishing up duties. He needed to contribute to that progress, notably as AI brokers started taking over work with actual operational and monetary penalties.

He weighed the choice fastidiously. His job at Optiver was prestigious, and the trail forward was clear. However AI was advancing shortly, and he judged that the chance to work on the entrance of the sector was value pursuing. He resigned with out one other job ready.

The choice to depart got here first. Discovering the proper downside to work on got here later. Whereas he searched, Tang got here throughout Fragment, an utilized AI startup in San Francisco, and the issue it was engaged on. Procurement is the method corporations use to purchase the products and supplies they want. It underpins the bodily economic system, but a lot of it’s nonetheless dealt with by hand, by repetitive workflows that eat up time.

That appeared to him a pure place for AI brokers to assist. However automating procurement requires greater than an agent that may full a activity beneath ultimate situations. The system should deal with difficult workflows, reply appropriately to sudden conditions, and make selections that companies can depend on. In procurement, a mistake can value actual cash, so reliability is a requirement slightly than an additional advantage.

Fragment was already rising shortly, constructing AI brokers for enterprise procurement workflows, and has since raised a $45 million Sequence A. Tang joined as a founding engineer, taking over work that may assist decide whether or not its brokers might function reliably in actual enterprise environments.

At the moment, Tang works as a founding engineer at Fragment, main efforts to guage the reliability of its AI brokers earlier than they’re deployed in actual enterprise environments. The central query is easy to state and onerous to reply: can you understand whether or not an agent is dependable earlier than it acts?

Testing an agent in a managed demonstration is just not sufficient to determine that it may possibly deal with the calls for of an actual enterprise. Manufacturing environments comprise complicated knowledge, uncommon conditions, and selections whose penalties might not be apparent till one thing goes fallacious. An agent that performs properly on a slender set of duties should still fail when situations change.

Tang’s workforce addresses this by testing brokers in simulated environments constructed from actual manufacturing knowledge. These simulations recreate life like working situations, permitting the workforce to look at how brokers behave throughout a spread of conditions earlier than deploying them to prospects. The purpose is to uncover weaknesses early and perceive the place a system might fail, slightly than uncover these failures after it has begun dealing with actual procurement work.

The usage of manufacturing knowledge is essential as a result of it helps make the simulations consultant of the environments through which the brokers should function. The workforce can check how techniques reply to the complexity of precise enterprise workflows, together with conditions that won’t have been lined throughout coaching. This makes analysis greater than a last test earlier than launch. It turns into a method to measure reliability, determine limitations, and enhance the system earlier than prospects rely on it.

His view is that how an agent is evaluated issues as a lot as how succesful it’s. A robust system that can not be trusted is not going to be given actual items or actual cash to deal with. For companies contemplating AI brokers, the flexibility to display reliability is crucial to shifting from experimentation to sensible use.

AI techniques engineering can also be a younger self-discipline. In older fields, it takes a few years to achieve the frontier. Right here the frontier is just a few years previous, which suggests the individuals constructing manufacturing techniques right now are serving to resolve how the work is completed. Tang says he desires to be one in every of them. His private purpose is to grow to be a world-class AI engineer and to assist set requirements for a way these techniques behave.

For now, his focus is on serving to Fragment develop into a number one platform for autonomous procurement. As AI brokers tackle extra consequential enterprise duties, the flexibility to check them in life like situations will stay essential to their adoption. Fragment’s work addresses that problem by making it attainable to guage brokers in opposition to the complexity of actual procurement operations earlier than prospects depend on them.

Fragment has simply come out of stealth with $56 million in complete funding, in keeping with reporting from Axios and AI Journal. The Sequence A was led by 8VC, whereas Normal Catalyst and A* co-led the seed spherical. Roelof Botha, former head of Sequoia, can also be an investor. The funding will help continued growth of Fragment’s know-how, growth throughout source-to-pay and deployments with extra giant enterprises.

Long term, he sees procurement as a place to begin. He photos the identical strategy spreading throughout the bodily items trade, altering how merchandise are sourced, made and moved world wide. Getting there, in his view, comes all the way down to constructing brokers that companies are prepared to belief with actual items and actual cash.