Written by Shen Haoming.
From Prompts to Systems
5 min read · Part of Sponsoring students at AIE became a test of Singapore's AI ambition
Written by Shen Haoming.
5 min read · Part of Sponsoring students at AIE became a test of Singapore's AI ambition
When Singapore’s Minister for Foreign Affairs, Dr Vivian Balakrishnan, walked onto the stage at AI Engineer Singapore and revealed that he had been working on his own AI “second brain” for the last few months, the room immediately understood that this was not going to be a typical tech conference keynote. Instead of talking about national AI strategies, he spoke like a builder. He explained how he used containerised AI agents, memory systems, local LLMs, and workflow automation to handle his daily workflow. More importantly, he reminded the audience that while AI can outsource computation, memory, and drafting, it cannot outsource human understanding and accountability.
That moment captured the spirit of AI Engineer Singapore 2026.
Held from 15 to 17 May 2026, AI Engineer Singapore was the first AI Engineer conference in Asia, bringing together hundreds of engineers, founders, researchers, designers, and students to discuss what it actually means to build with AI today. Unlike many AI events dominated by marketing demos and futuristic speculation, this conference focused heavily on practical systems, production workflows, infrastructure, evaluation, security, and real-world deployment. As one of the 20 students selected by 65labs to attend through a sponsored scholarship, I expected to learn about new AI tools and model capabilities. What surprised me was how much the conference focused on turning AI into reliable systems that can actually work in production.
One of the themes throughout the event was that AI engineering is moving beyond simple code generation. The conversations were no longer about whether AI can autocomplete code. That question already feels settled. Instead, builders were discussing how AI agents can participate across the entire software development lifecycle, planning, implementation, testing, debugging, deployment, observability, and maintenance. In Gavriel Cohen’s talk, NanoClaw demonstrated how autonomous agents could support pull request triaging, code review, and testing. Thibault Sottiaux’s discussion around Codex also reinforced the same idea, the real engineering challenge is no longer just generating code, but building the harness around the model, including memory, context management, sandboxing, permissions, tooling integration, and safe execution.
Similarly, another moment that stood out to me was when I asked both the Cursor team and the Codex team what made their AI coding tools different from the many other tools in the market. Interestingly, both pointed to the same idea, the harness. In other words, the advantage is not only the model itself, but the surrounding environment that lets the model work effectively, the way it receives context, interacts with files, runs commands, uses tools, reviews changes, and stays within a controlled workflow. That suggests the future of AI coding will not be won by model intelligence alone, but by the engineering systems that wrap around the model and turn it into a reliable software teammate.
To make AI agents reliable in real engineering environments, builders now need to design entire operational layers around the model itself. This includes long-context memory systems that allow agents to retain relevant information across sessions, sandboxed execution environments that safely isolate commands and generated code, and orchestration frameworks that coordinate multiple specialised agents together.
Equally important is the idea of keeping credentials away from the agent runtime. The agent side should be treated almost like an untrusted environment. If an agent can reason, execute commands, access tools, and modify files, then giving it direct access to production credentials creates serious risk. A safer design is to use zero-credential agents, where sensitive permissions are handled through controlled approval layers and isolated environments. When a human approves a request, only the specific tool action should be executed, not the agent itself. This keeps the agent useful while preventing it from becoming the direct holder of sensitive access.
Beyond the technical details, the bigger message was clear, AI transformation is not happening through massive top-down replacement of entire industries. It is happening workflow by workflow, team by team, and person by person. Many speakers framed AI not as a replacement for engineers, designers, or operators, but as infrastructure that can augment human work when it is designed with the right safeguards.
That idea may ultimately become one of the defining characteristics of Singapore’s AI ecosystem.
Rather than competing directly with frontier AI labs on massive model training, Singapore appears increasingly positioned to become a hub for applied AI engineering, integrating models into practical systems, enterprise workflows, government services, and operational infrastructure.
A strong AI engineer today is no longer simply someone who understands machine learning models. They must also understand system architecture, observability, user experience, security boundaries, evaluation methodologies, and human-computer interaction. The most effective teams at the conference were rarely isolated AI researchers. They were cross-functional groups combining infrastructure engineers, designers, product thinkers, and operators. The barrier to entry for building meaningful AI systems has dropped dramatically. The tools are increasingly accessible. Open-source ecosystems are evolving rapidly. Small teams and even students can now prototype systems that would have required enormous resources just a few years ago.
At the same time, AI Engineer Singapore also emphasised that responsibility matters more than ever. As AI systems become more autonomous, builders must think carefully about reliability, transparency, accountability, and human oversight. The most impressive talks were not the ones promising fully autonomous replacement of humans. They were the ones focused on augmenting human capability responsibly.
In many ways, AI Engineer Singapore felt less like a conference about artificial intelligence and more like a conference about human amplification.
The message I left with was simple: the future will belong not only to those who build powerful models, but to those who understand how to integrate AI meaningfully into real human workflows.
And judging from the conversations happening in Singapore right now, the next generation of AI builders is already starting to do exactly that.