How Grab is reinventing its engineering culture around the speed of AI

Suthen Thomas Paradatheth, Grab’s chief technology officer, argues that while artificial intelligence will make code cheaper to write, it will also make code checks rarer.

This change requires leaders to reimagine operations, employment, and physical operations in Southeast Asia before speed becomes an issue.

The hidden costs of building software quickly

As Grab releases new software tools at a record pace, leaders now need to understand which numbers are indicative of a better business, not just a busy staff.

Old ways of measuring work can push companies toward the wrong goals. This means management can no longer rely on traditional metrics to track actual results.

“90% of our engineers use some form of AI coding assistance every day,” Paradatheth said. “We didn’t force anything. We made the tool available, taught people the skills on how to use it effectively, and then set it free.”

Production increases, but it becomes difficult to measure the amount of work done

“Engineering productivity always comes with a lot of caveats and asterisks,” he reports.

Still, the numbers speak for themselves. By using merge requests as proxies, Grab increased production per person by approximately 40% and reduced the time required for similar sized tasks by 20-30%.

He adds that engineering achievements are just one piece of the puzzle, noting that true AI effectiveness requires company-wide organizational change in parallel to individual improvement.

Changing team roles and the risks of easy coding

Enabling more people to build software speeds up the work, but creates new hiring and oversight issues. “Software engineering fundamentals are still important,” Paradates says.

When it comes to employment, he draws the line. “You can’t come in here and say, ‘I can ask any agent to do this, but I don’t know what that agent did.’ We also want the fluency of the AI ​​and the sense of ownership to act like an owner instead of waiting for instructions.”

Lines between departments blur as non-technical staff can automate their work

Our legal team built an automated tool that reduces the initial NDA review from hours to minutes. Meanwhile, the design team created a similar tool called Mosaic, which generates brand-specific illustrations in a fraction of the usual time.

“When a tool is deployed into production and exposed to end customers, it must be reviewed by an operations engineer,” he explains. “If we use it internally, we don’t want the engineering department to be the gatekeeper.”

Make humans responsible for independent systems

Removing internal gatekeepers speeds up development, but shifts the pressure elsewhere. Weak checking mechanisms risk turning sophisticated engineering achievements into vulnerable systems.

To address this, Grab has built its workflow around four operational steps.

• Change engineering work from direct prompting to task handoff, allowing the system to operate on its own.
• Organize business and technical data so independent systems can get context without crashing
• Extend your checking processes with automated tests to prevent software bugs and verify unexpected behavior
• Use quadratic models to check output, monitor automated decisions, and assign ultimate responsibility to specific humans.

Software creation changed when systems no longer waited for human input.

This workflow changes the way software is generated. Engineers can assign tasks to agents, walk away, and come back later to review the generated code, increasing speed through asynchronous workflows.

“Code can be generated in large quantities, and the new bottleneck is the inability of humans to review all the code that is created,” Paradatheth warns. “Ultimately, you’re responsible for getting it all the way to production. So how do you review it?”

To manage this, the company has invested heavily in “harness engineering” to keep the codebase easy for AI to read while scaling automated tests to catch errors before they reach production.

“The chain of responsibility doesn’t end with the AI, it ends with the human,” he says. “All leaders in my organization are expected to use agent engineering to make changes to the production environment.”

Build rules directly into your software

Holding humans accountable for checking code requires a managed platform, and leaders face the challenge of providing computing power to their staff while maintaining oversight.

Grab’s answer is GrabGPT, born from a failed internal chatbot. When AI excitement grew in 2022 and 2023, external tools had obvious security risks, so infrastructure teams built secure internal interfaces instead.

“Despite the name, GrabGPT does not use just one vendor,” says Paradatheth. “You can have closed models like Gemini, Claude, GPT, and open weight models like Qwen.”

GrabGPT also acts as a router and abstraction layer with audit logging, controlled onboarding, usage metering, and cost controls.

Find hidden software issues faster

This centralized router shifts the focus of security from individual tool selection to the operation of the entire system, allowing leaders to prepare for an environment where structural weaknesses can grow more quickly.

“A useful mental model is that AI is the amplifier of everything,” Paradates says. “If you have good software engineering practices, that’s amplified. If you have a potential risk in your system, that risk is amplified.”

Unpredictable system behavior can expose previously hidden weaknesses

“You have to start thinking in terms of potential risks,” Paradates warns. “Vulnerabilities may have always existed, but now we have agents that behave in a non-deterministic manner. You can’t just ship your code to production and hope for the best.”

But those same forces can also work in reverse, with automated reviews now able to find and fix weaknesses much faster than human reviewers alone.

Correct slight delays in physical operations

This bandwidth for risk assessment extends to the physical logistics networks that connect businesses and consumers, and robotics is seen not as a replacement for human workers but as a tool to eliminate wasted operational time.

“The robot can be loaded at the counter, meet the driver at the curb, deliver straight to the door on the other side, and handle the first and last few meters of the trip,” says Paradates.

These walking stages consume about 10 percent of a driver’s time, so removing them allows couriers to process more orders and increase revenue.

Adapt autonomous systems to different streets

Robotic enhancements can address walking delays, but achieving full self-driving capabilities will require navigating economic and regulatory conditions that vary widely from city to city.

“It’s going to be a long road for self-driving cars before it becomes more widespread in Southeast Asia,” Paradates explained.

The road ahead is complex. “Challenges include unit economics and adapting to diverse situations. Cities are full of motorbikes and bicycles, and there are often no bike lanes.”

When it comes to autonomous delivery robots, the company builds everything in-house. For passenger vehicles, smooth integration into existing markets requires a multi-partner approach. asian technology