Not so super: AI agents have limits too

Welcome to Tech in Asia’s daily newsletter, your essential dose of Asia’s tech and startup buzz. Not on the mailing list? Register here. Got a story tip? Send it to editors@techinasia.com.

IN FOCUS

• How justified is the hype surrounding AI firm Manus? We put it to the test
• A new breed of VCs for AI startups
• Implications of Meta-Scale deal on AI data startups

---

Welcome to The Prompt! Delivered every Monday, this free newsletter breaks down the biggest stories and trends in AI. Get all our content, including The Prompt, by registering for a Tech in Asia account. Alternatively, sign up here if you just want to receive this newsletter.

Hi there,

AI agents were the talk of the town at last week’s SuperAI event in Singapore. There, companies like Manus AI and Auki Labs took the stage and shared how they are lining up innovations that will drive the tech industry’s next big thing.

While terms like “agentic AI” might sound like the latest buzzword, the idea behind them is simple: We’re no longer just asking AI to generate content, we’re starting to delegate full tasks, decisions, and workflows to these autonomous systems.

But just like in the early days of generative AI, the road ahead won’t be smooth. Hallucinations still plague AI models, and for agents, even something as mundane as a CAPTCHA can be a roadblock, as my colleague Scott explores in this week’s first Big Story on AI firm Manus.

That’s because today’s internet was designed for humans, not machines trying to navigate on our behalf. Getting these agents to communicate better with existing digital infrastructure is our biggest task if we ever want them to take real work off our plates.

While that remains outside the control of one company alone, some are trying to make their agents more efficient in the meantime. Singapore-based SapiensAI, for example, has developed its own language to communicate between agents within its recently launched application, Agnes AI.

Rather than using natural language, agents within Agnes communicate with one another using a “code-like” language that the company says helps the AI execute tasks faster, at higher accuracy, and with 40% to 70% fewer tokens.

That same builder mindset is also showing up on the investor side. In our second Big Story, we look at how a new wave of VCs is writing code – not just checks – to stay competitive in the AI race.

But efficiency is just one part of what shapes AI performance, as the quality and control of training data matters just as much. In this week’s Hot Take, I look at Meta’s surprising new investment deal with Scale AI and why some of the AI firm’s top clients are now backing away.

On that note, this newsletter will now land in your inbox weekly, because, as we all know, AI doesn’t wait.

---

THE BIG STORIES

1️⃣ Manus shines at research but crumbles against CAPTCHAs

---

THE HOT TAKE

---

DEEP READS

---

Connect with Kevin Aluwi, Alex Hungate, Gita Wirjawan, Hans Patuwo, and more!

---