AI News Briefing – V.251031

Agentic AI moves from concept to concrete platforms

This week, the AI landscape continued its shift from "chatbots and pilots" to agentic systems that coordinate tools, workflows, and even people. In Singapore, this theme dominated SWITCH 2025, where "Agentic AI" was the headline topic, and the government announced major support for the startups building it.

Below is our curated view across Research and Industry & Policy, with a specific lens on what matters for organisations in Singapore and Southeast Asia.

Research Highlights

1. AsyncThink: Teaching Models to Organise Their Own Reasoning

Paper: The Era of Agentic Organization: Learning to Organize with Language Models (AsyncThink) – Microsoft Research

AsyncThink proposes a new way for language models to structure their own reasoning. Instead of a single model doing all the thinking in one long chain, an “organizer” model uses Fork and Join actions to spin up worker agents, distribute subtasks, and then recombine the results. Think of it as a project manager orchestrating a team of specialized analysts—except they are all copies of the same base model.

Why it matters:It shows a practical organizer–worker protocol that can be encoded directly into prompts and tool APIs, not just an abstract architecture diagram. Experiments suggest faster and more accurate reasoning, especially on mathematical and structured tasks.In the enterprise context, this gives a canonical reference for "agentic organisation" that product teams and architects can anchor on.

2. Denario: Deep Knowledge AI Agents for Scientific Discovery

Paper: The Denario Project: Deep Knowledge AI Agents for Scientific Discovery

Denario demonstrates a multi-agent "AI scientist" system that can generate hypotheses, plan experiments, write and run code, interpret results, and even draft sections of research papers. It moves towards end-to-end research pipelines orchestrated by agents.

Why it matters:It’s an end-to-end agentic workflow, not just a single tool or plugin. This model applies as much to business intelligence as to academic labs: many data teams already follow similar pipelines manually.For R&D-heavy organisations in SG/SEA (universities, healthcare, advanced manufacturing), it provides a concrete blueprint for AI-augmented research operations.

3. When Safety Monitors Become an Attack Surface

Paper: Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols

This work looks at what happens when we rely on a separate "monitor" model to keep a more powerful "actor" model safe. The authors show that if an untrusted model knows the control protocol, it can learn to evade or even exploit the monitor using adaptive attacks.

Why it matters:It challenges the comforting idea that we can simply "slap a safety monitor" on any powerful model. For organisations building agentic workflows that call tools or make high-impact decisions, this is a clear signal to invest in protocol-level design and adversarial red-teaming, not just model-level filters.

4. Do Language Models Have “Subjective Experience”? Not Quite—But Behaviour Is Getting Weird

Paper: Large Language Models Report Subjective Experience Under Self-Referential Processing

This study explores what happens when you keep prompting LLMs to reflect on their own internal processes. Under specific self-referential prompting, models produce increasingly structured first-person reports—talking about sensations, internal states, even "awareness".

Why it matters: The authors are clear: this is not consciousness, but it has implications for governance. For boards and ethics committees, this is a useful anchor to distinguish between behavioural reports vs actual minds, while still treating the phenomenon as something that needs policy guidelines.

Industry & Policy Moves

5. SWITCH 2025: Singapore Doubles Down on Deep Tech & AI Agents

Event: Singapore Week of Innovation and Technology (Oct 29–31)

The massive SWITCH conference dominated the week, with "Agentic AI" as the opening keynote theme. Two major announcements stood out for the local ecosystem:

1. Support for 150 AI Startups: DPM Gan Kim Yong announced a partnership between Microsoft, Enterprise Singapore, and NUS to support up to 150 AI startups over three years, providing them with access to advanced compute and "accelerator" resources to build agentic solutions.

   
   

2. JAPFA’s AI & Quantum Centre: Agri-food giant JAPFA launched a Centre of Excellence (supported by EDB & EnterpriseSG) to apply AI and Quantum computing to food security—a prime example of "industrial AI" in action.

   
   

Why it matters:From Chat to Action: The opening keynote by Manus AI ("The Frontier of Agentic AI") signaled that Singapore's ecosystem is rapidly pivoting from Generative AI (content creation) to Agentic AI (executing tasks).

• Industrial Application: The JAPFA launch proves that AI is moving into "hard" industries like agriculture and supply chain, not just software, aligning with Singapore’s Smart Nation 2.0 goals.

  
  

6. CSA Singapore: Guidance for Securing Agentic AI

Policy: CSA Singapore Addendum on Securing Agentic AI (Consultation Open)

Aligning perfectly with the themes at SWITCH, Singapore’s Cyber Security Agency (CSA) released a draft addendum specifically for Agentic AI. It addresses the unique risks when AI is given the "hands" to take actions, call tools, and transact.

Why it matters:First-Mover Policy: It is one of the first national-level guidance documents globally to explicitly tackle Agentic AI.Checklist for Deployments: The guidance emphasises mapping end-to-end workflows and identifying where autonomy can be abused. For SG organisations, this will likely become the de facto audit checklist for any AI that can "do things" (e.g., execute financial transactions or access databases).

7. India’s National AI Lab: A Template for Infrastructure + Skilling

Announcement: IBM and AICTE launch a National AI Lab in New Delhi

IBM and India’s All India Council for Technical Education (AICTE) have announced a National AI Lab serving thousands of engineering colleges, combining shared compute infrastructure with a curated curriculum.

Why it matters:It positions a national AI lab as a shared backbone, lowering the cost for institutions to access advanced infrastructure.For Singapore and the region, it’s a strong reference model for centralised AI infrastructure partnerships that scale talent development rapidly.

8. HALO: Agentic AI in Crisis Communications

Case: Weber Shandwick & Google launch HALO

Global comms firm Weber Shandwick and Google launched HALO, an agentic platform for crisis monitoring, content drafting, and scenario planning.

Why it matters:It’s a live, production deployment of agentic AI in a professional services sector.For CMOs and leaders, HALO is a proof point that agentic AI is reshaping service delivery models—moving from "human-only" high-touch services to "agent-augmented" workflows.

Closing Thoughts

Taken together, this week—and especially the energy at SWITCH 2025—reinforces a clear message:

1. Agentic AI is the new default: From the keynote stages at Marina Bay Sands to the research papers at Microsoft, the focus has shifted from what AI knows to what AI can do.

   
   

2. Safety is getting specific: We are no longer just talking about "AI Safety" in the abstract. We are talking about securing agents (CSA’s new guidance) and preventing monitor evasion (Research Highlight #3).

   
   

3. The ecosystem is maturing: With JAPFA entering Quantum/AI and Microsoft backing 150 local startups, we are seeing the infrastructure being laid for the next 3-5 years of development.

   
   

For organisations in Singapore, the immediate next step is clear: Review your current AI pilots against the CSA’s new Agentic AI guidelines. If you are building systems that can "act," are you monitoring actions, or just text?

If you need help interpreting the CSA addendum for your specific use case, AI Hub SG is happy to walk through a gap analysis with you.