Does every task need an AI agent? No. Clear, repeatable steps may call for a workflow. Consider an agent when the goal is clear but the route may change. Start simple, set boundaries and check the result. Full article: medium.com/@arthur.lin.ax/ag… #AI #Agent101

Singapore’s manufacturing AI push is gaining momentum. More than S$45 million has been set aside over the next three years to help manufacturers adopt AI. That wider push raises a practical question for industry leaders: how can AI be put to work in live operations while keeping decisions traceable, people accountable and data under the client’s control? At FIX 2026 and Plug and Play Korea Expo in Daegu, physical AI was on display across robotaxis, delivery vehicles, humanoids and factory robots. For industry, the value grows when machines, sensors and people share a live model of the site, with a named owner for every decision. On the Plug and Play Korea main stage, 2MR Labs CEO Arthur Lin presented Giants OS: a live model of an operation that AI agents can act on, with every action awaiting approval from the person responsible for the decision. At booth 79, Cody Yu, Head of Product and Research, and the team ran live demonstrations of a factory twin, a pipeline corridor and the Giants OS workspace. At Runhai Industrial, a textile plant in Shantou, the team built a model from two phone clips totalling 149 seconds, with 4,413 time-stamped evidence frames. In the pilot, agents run the plant’s line controls with human approval. Runhai is a paying client, and its data stays on its own in-country infrastructure. 2MR Labs is a Singapore-based company, building from Singapore to support organisations here and work with partners across the region. We welcome conversations with manufacturers, investors and public-sector leaders exploring where physical AI can create measurable value.

What’s actually inside an AI agent? Five layers sit behind every agent you’ve seen demoed. Here’s how they fit together, and why reliability matters. First, a definition. An AI agent takes a goal, plans the steps, uses tools to carry them out, and checks the result. A language model is at its centre. Most of the work is the system built around it. Does every task need an agent? Anthropic draws a useful line: workflows follow predefined code paths; agents let the model direct its own steps and tool use. Start with the simplest option that works. Agents fit when the path can't be written down in advance. Layer 2 is the heart: the runtime loop. Observe → Think → Plan → Act → Reflect, repeated until the task is done. This pattern traces back to ReAct (2022), which interleaved reasoning with actions. Layer 1, Inputs & Context: the request, conversation history, files and outside signals. Here’s how they fit together, and why reliability matters. Layer 3, Intelligence Core: model routing, reasoning and prompt templates. The runtime moves back and forth between the two on every step. Layer 4, Capabilities & Tools. Function calling lets the model trigger structured actions. MCP (Model Context Protocol) gives tools and data sources one standard way to connect. Execution happens in a sandbox. Layer 5, Memory & Infrastructure: short-term context, long-term memory, vector search and knowledge bases, plus the logs and metrics that show what the agent did. Why the loop needs checks: if each step is 95% reliable, a 10-step task succeeds about 60% of the time (0.95^10). At 99% per step, 20 steps still lands near 82%. Reflection, verification and human checkpoints exist to break that chain. Wrapped around everything: governance. Access control decides what the agent may touch. Evaluation tests its outputs. Feedback from real results improves prompts, tools and behaviour over time. Reliable agents are built on memory, tools and feedback. Full breakdown with the end-to-end flow: medium.com/@arthur.lin.ax/th… Next in the series: the main agent architecture types.

Meet Arthur Lin @ArthurlinSG — Founder of 2MR Labs. 12 years in Special Forces. Two combat tours in Afghanistan. Builder of Singapore's Special Operations doctrine. "This is no longer just an opportunity — it is a calling." Now building sovereign AI infrastructure for governments that can't afford to get security wrong. Full story: medium.com/@2MR_Labs/arthur-…

We're reintroducing 2MR Labs to the world today. 2MR = Tomorrow Labs. We started with AI for complex information environments. The world told us what came next. Geopolitical instability is structural. Governments need AI infrastructure that is sovereign, secure, and built for real operational environments. That is what we are building. Full article: medium.com/@2MR_Labs/2mr-mea…

🎄🎅 2MR Labs wishes you a Merry Christmas! 🎁 Thank you for being a part of our exciting journey thus far. Let's spread kindness to all. 🤗✨, and have yourself a joyful holiday season.

2MR Labs Frontier Tech Seminar is here! 🎙️: Arthur Lin, 2MR Labs Eugene Lim, The Assembly Place Clarence Guo, Jacque Law LLC Chyi Yan Hshieh, Polygon Labs Savitha Ramasamy, A*STAR Fan Zhang, Zhang Fan AI Foundation Fangda Wan, NetEase Games Charlie Hu, CommonWealth Capital

For more details on the event, check the post below

Why choose 2MR Labs? ✅ Customized end-to-end solutions ✅ Community building experience ✅ Tap into our wide Web2 & Web3 networks