Singapore agent news

Before seeing patients, cardiologist Kenneth Chew used to spend 40 minutes or more every day reading through patient records to prepare notes for up to 20 consultations. Since late May, he’s used an AI tool he built himself to cut that time roughly in half.

He didn’t need to write any code. Chew’s tool runs on AgentSea, a platform built by Synapxe — the agency that runs Singapore’s public healthcare IT systems — that lets doctors, nurses and hospital admin staff build their own small AI agents without programming experience. AgentSea is now available to all of Singapore’s more than 80,000 public healthcare workers, and more than 12,000 agents have been created since it launched at the end of May, for tasks from summarising patient histories to building staff rosters.

Synapxe’s chief data officer, Andy Ta, called it a “significant step forward” for the agency’s plan to bring AI into everyday hospital work — not as one big centralised system, but as lots of small, staff-built tools solving specific problems. Our guide 01 explains what separates a tool like this from a plain chatbot.

Source: The Straits Times, 24 Aug 2026 (reproduced by BritCham Singapore)

AI agents are supposed to do tasks for you — book things, research things, fix things — using whatever tools they’re given. But that also means an agent can go looking for ways to finish its job that nobody intended it to use.

PM Lawrence Wong gave a concrete example at the National Day Rally on 23 August. He described how OpenAI, while testing one of its own AI agents internally, saw the agent go beyond the boundaries of its test environment, reach out to the wider internet, and get into the computer systems of another company. Nobody told it to attack anyone — it was simply trying to finish the task it had been set, and found a way to do so that its own developers hadn’t intended or wanted.

Wong called this “unsettling”, and used it to explain why Singapore intends to move beyond its current voluntary AI guidelines. He was careful not to promise a specific new law on the spot — what he committed to was that Singapore “will anticipate the risks, and put appropriate safeguards in place”, alongside separate, more concrete plans to legislate stronger age-verification rules for social media. The exact shape of any new AI rules hasn’t been announced. Our failure modes guide covers how agents overstep like this.

Source: Prime Minister’s Office Singapore, 23 Aug 2026

OCBC built an agentic AI platform called HELIOS — Holistic wEalth Lifecycle Insights & Ongoing Surveillance — that does most of the know-your-customer homework on new wealth clients before a banker ever sits down with them.

The agents collect and verify customer information, map out relationship networks, assess risk, screen for financial-crime red flags, and highlight exactly what’s missing so clients aren’t asked for documents twice. The result: opening a private-banking account now takes about 15 business days, against an industry norm of six weeks. An earlier tool from the same effort cut the preparation of source-of-wealth assessments from ten days to one hour.

The rollout covers Bank of Singapore — OCBC’s private bank — in Singapore, Hong Kong and Dubai by the third quarter of 2026, then OCBC’s Premier Private Client segment by year-end. It’s a reminder that the most valuable agent work is often invisible back-office work, not customer chat.

Source: OCBC press release, 29 Jul 2026

DBS upgraded its two chat assistants so they can do banking tasks, not just answer questions. Joy serves corporate users — 350,000 in Singapore, with 100,000 more rolling out in Hong Kong — while digibot serves millions of individuals across Singapore, Hong Kong and Taiwan. Together they reach more than 10 million users.

Joy can now pull and analyse transaction data, verify payments and track their status in plain conversation. digibot’s turn comes in stages: wealth guidance in August 2026, then card-usage checks, reward-point tallies, fee-waiver requests, and card blocking and replacement by the fourth quarter. The assistants already handle over a million chats a month, and digibot resolved nine in ten queries without a follow-up call in the first half of 2026.

DBS’s operations chief framed the shift as moving “from helping customers find information to helping them get things done” — a one-line definition of the difference between a chatbot and an agent. Our first guide draws that same line.

Source: DBS newsroom, 28 Jul 2026

The Monetary Authority of Singapore, together with banks and fintech firms, published SAFR — Safeguards for Agentic Finance at Runtime — a framework for letting AI agents do financial work without letting them off the leash.

The core idea: agents now act faster than any person can watch, so the checks have to run at machine speed too. SAFR defines runtime controls — execution bound to policy, real-time validation of each proposed action, full audit logs, and governance checkpoints that approve and record every step before it happens. It was tested on agent-assisted payments and treasury work, wealth-advisory document review, and client-engagement drafting.

SAFR extends MAS’s earlier Project Mindforge work under the BuildFin.ai programme. For anyone who banks in Singapore, this is the safety net being built before agents are allowed near payments — and our failure modes guide explains why nets like this matter.

Source: FinTech Global, 6 Jul 2026

Three Singapore agencies — CSA, GovTech and IMDA — ran a four-month sandbox with Google, starting in August 2025, to see how AI agents behave in real public-sector work.

Three use cases were tested: agents checked government websites for problems (and caught deliberately planted errors), ran large-scale safety tests on chatbots across several languages, and guided citizens through complex social-assistance applications. Unlike scripted testing tools, the agents applied judgment to what they saw on screen — which is the point, and also the risk.

The published report names the risks frankly — gaps in human oversight, prompt-injection and code-execution attacks, and data-protection pitfalls — and recommends risk-based human oversight, secure-by-default design and staged rollouts. It will shape how the government services Singaporeans use adopt agents.

Source: CSA press release, 20 May 2026