Grab seeks AI edge in product development using agentic engineering, vibe coding

[SINGAPORE] Artificial intelligence is aiding Grab in its product development, boosting productivity for engineers and allowing non-technical staff to launch prototypes for testing faster.

Take, for example, one of the products launched at the super app’s recent GrabX product showcase in April: Grab’s Virtual Store Manager.

It can take closed-circuit television feeds to “see” and “understand” what is going on, with a dashboard based on certain parameters.

This means there is no need for someone to physically review all the footage. A manager or management team can oversee an entire multi-store network from a single screen.

This product started life as a vibe-coded prototype. Vibe coding utilises AI to generate applications via prompts rather than having someone write the code. This helps those who have limited programming experience, allowing non-coders to build apps.

Virtual Store Manager’s creator, Muhammad Hanif Naufal Eka Wiratama, is not a software engineer, but was able to translate his idea into a prototype quickly without having to go back and forth with the engineering team.

According to Grab chief technology officer Suthen Thomas, this is where vibe coding has had the biggest impact on the product development process.

“We use it the way a designer uses a high-fidelity mock-up to quickly get something in users’ hands and observe how they actually behave, rather than asking them what they want,” Thomas told The Business Times.

A high-fidelity prototype is a version close to the final product, complete with visual design details and content.

SEE ALSO

A*Star, JTC, Grab enter partnerships with robotics firm Sharpa to build Singapore’s physical AI sector

Instead of reducing the need for humans, Grab’s AI push might result in the opposite

Grab unveils 13 AI-powered solutions for businesses, consumers, travellers

Still, Thomas is quick to add that while a product or project might have started out as a vibe-coded prototype, this serves only as a starting point for a problem, rather than for engineering work.

For software engineering work, more thought and care needs to be put into the code, from how it interacts with other back-end systems to how its logic works. A vibe-coded prototype would not need to take into account such considerations when coded by AI.

Thomas likens vibe coding to “hot-wiring a car where it gets you moving, but you wouldn’t drive it at scale”.

Building a product or feature to last beyond the initial testing phase requires engineering rigour. Here, context, architecture, testing and verification are relevant for the code for scalability and maintenance.

“We maintain a strict boundary between casual prototyping and actual production to ensure our engineering standards remain uncompromising,” said Thomas.

Agentic engineering and the drive towards cyborganisation

On the engineering front, Grab wants AI agents and humans to work together in a product development approach known as agentic engineering.

This is part of Grab’s vision of becoming a “cyborganisation”, where humans and AI agents work together seamlessly to drive value, said Thomas.

It comes from the science fiction concept of a cyborg, in which technical and organic components are fused together to create a stronger and more capable entity.

With agentic engineering, tasks are delegated to AI agents to execute before humans review the output, rather than for humans to code together with the AI assistant.

“Today, the engineer is able to assign higher-level tasks or goals to the AI agent, which then might take minutes or hours to complete,” said Thomas. “The human only steps in at the end to check the finished work.”

Not a straightforward process

Thomas outlined a pause in business-as-usual activities in 2024, when Grab undertook its first generative AI upskilling and experimentation exercise across the entire company.

The nine-week session was part of a move to invest heavily in the company’s talent, sustaining this momentum with hands-on experiences, internal workshops and community-led learning of AI.

It paid off, with more than 90 per cent of engineers using AI coding tools daily.

Productivity is on the rise as well, with about 40 per cent more merge requests per engineer – this refers to the request for the code to be integrated into an application, meaning that applications are completed faster – and a reduced 20 to 30 per cent development time.

As for the non-engineering teams, Thomas said that they are also not left behind, as more are using AI to automate tasks and workflows.

“In the long term, this will multiply our engineering bandwidth to create more possibilities for our users and partners across South-east Asia,” he said.

AI will speed up work

AI is not a bolted-on feature, but a foundation on which Grab is building its products and features, Thomas added.

Data from over 20 billion rides and orders, along with real-time signals, gives Grab the edge in sensing, predicting and responding to the complexities of operating in South-east Asia, he explained.

There are some challenges in utilising AI at work, as he warned against the normalisation of deviance – a concept coined after the Challenger space shuttle disaster. This refers to a deviance from the rule becoming culturally normalised.

In engineering, as AI-assisted work ramps up in speed and volume, Thomas is watching any potential erosion of engineering standards. This is even more important, given the scale that Grab is operating at in South-east Asia.

“The faster you can build, the more important the discipline around what you accept,” he added.