British robotics champion now leading enterprise revolution in autonomous AI agents

Photo courtesy of Edward Upton.

Opinions expressed by Digital Journal contributors are their own.

Edward Upton stood before 5,000 developers last month at the MLOps Community conference, dissecting why autonomous browser agents fail in production. His presentation, titled “Catastrophic Agent Failure and How to Avoid It,” drew from direct experience building systems that book healthcare appointments and process insurance claims—tasks where mistakes carry consequences. “Customers expect human-level accuracy from agents,” Upton explained during his talk. “They want systems that adapt to changes and understand context, just like a person would.”

The 27-year-old founding engineer at Asteroid has spent the past year constructing technical infrastructure that insurance companies and healthcare providers trust with sensitive operations. His team’s ProductHunt debut in August captured second place on launch day and third for the week, signaling demand for browser automation that doesn’t break under real-world conditions.

Building agents that don’t collapse under pressure

Upton’s path to enterprise software began in Wales, where his robotics team won the national final in 2017, before claiming the best robot design award at the UK and Ireland championship. The squad represented Wales at the global finals in St. Louis, finishing 44th among international competitors. Later, his sixth-form team won the PA Consulting Raspberry Pi Competition, demonstrating early ability to deliver complete technical solutions.

He graduated from the University of Warwick with first-class honors in Computer Science, earning recognition as the best graduating MEng student in his department. His solo dissertation developed a LEGO brick identification mobile application, while his team project built a customizable streaming platform—both received top marks. The University of Warwick awarded him its highest departmental distinction for consistent technical performance across four years of study.

Upton spent two years at Netcraft following a summer internship, building services to combat online fraud for major banks and brands. The work taught him to construct systems that handle sensitive data in regulated environments. That experience directly informs his current role, where clients in healthcare and insurance require guarantees that automated agents won’t make irreversible errors.

Graph architecture replaces single-path execution

The asteroid’s technical foundation reflects the choices Upton drove from concept through implementation. The platform employs graph-based agent construction, rather than the single-path execution models commonly used by competitors. Each node represents a specialized operation with defined parameters for when to follow deterministic scripts versus when to engage generative AI responses.

The architecture addresses a persistent problem in browser automation: context windows that expand as agents progress through complex tasks, driving up inference costs and latency. Upton’s system implements smart context pruning and subtask isolation. Each agent component focuses on specific objectives without carrying unnecessary conversational history. The hybrid execution model caches deterministic scripts for repeatable operations but switches to AI-driven responses when unexpected conditions emerge.

His work includes a hybrid approach that blends DOM-based precision with vision-based adaptability. Document Object Model parsing offers exact element targeting when webpage structures remain stable. Computer vision handles situations where layouts shift or dynamic content renders traditional selectors ineffective. The switching mechanism operates automatically, creating resilience without sacrificing speed.

From academic excellence to production systems

During his MLOps presentation, Upton detailed a healthcare appointment booking incident where an agent reached what appeared to be a successful completion state but had actually failed to book the required appointment. The case illustrated why flexible evaluation systems matter. Different customers need different success criteria. Reaching a specific page may signal success for one client but failure for another, particularly in quality assurance testing scenarios.

Asteroid’s platform now enables non-technical users to create browser agents through a redesigned, graph-based interface. The company completed Y Combinator’s winter cohort earlier this year and released its self-serve platform publicly in August. Dozens of customers have used the system to build agents that automate data entry into electronic health records, carrier portals, and supply chain management interfaces.

Upton’s conference talk concluded with technical recommendations drawn from production deployments: scope what agents can access, build in failure points rather than biasing toward progression, and implement evaluators that run as discrete systems after execution completes. These aren’t theoretical principles—they’re lessons extracted from systems handling real patient data and insurance transactions.