Agents live where your team already is.
One agent, dozens of channels, no new app to install. In Malaysia WhatsApp reaches 90.7 percent of internet users, so the interface barrier is zero. This is why adoption is bottom-up, and why it outruns IT.
The adoption evidence →
Agents plug into the tools you already run.
Ordinary APIs and MCP connectors turn a model into a working teammate across finance, ticketing, design, data, CRM and files. The agent is only as useful, and only as dangerous, as the tools you hand it.
How the agent loop works →
Agent Nara
OpenClaw runtime on a DeepSeek V4 Pro and Flash mix, roughly RM 120 a month at the stated usage assumption. Sandboxed, querying public Malaysian data only. Replies take 30 to 60 seconds.
Message Nara on WhatsApp Demo guide and starter prompts →
Agent Caoimhe
A custom agent on OpenAI GPT 5.6 Sol, about RM 2,050 a month at premium usage. Ask Caoimhe and Nara the same question and compare depth, caution and speed. The cost gap is the management lesson.
Message Caoimhe on WhatsApp Demo guide and starter prompts →
The difference that matters.
Chat apps converse but cannot act, and remember nothing across tools. Agents act, remember, and run around the clock. Your junior analyst still wins on judgment, exceptions and empathy. The play is agents for volume, humans for judgment.
What this does to research work →
Now the slide the vendors skip.
Prompt injection, credential exposure, shadow agents and missing audit trails, all documented with dates and sources. None of it argues for a ban. It argues for governance, because a ban keeps the agents invisible.
The downside, documented →
What a governed agent looks like.
Six rules: sandboxed, least privilege, approval gates, logged, right-sized models, and data stays home. Nara and Caoimhe run under all six, which is why we handed them to a room full of strangers.
The six rules and the ninety days →
What you actually need to build one.
A small cloud server or a spare laptop, OpenClaw plus the model key of your choice, a dedicated SIM that is not a personal line, a spare phone to pair WhatsApp, the tools your agent needs, then the lockdown. Building is a weekend. Governing is the real work.
How to govern it →
Advanced deployment.
Multi-model setups, local models on your own GPUs, custom frameworks beyond OpenClaw and Hermes, longer self-improving loops, computer and mobile control, and multi-agent teams. Every step up in capability raises the governance bar.
The governance playbook →
Agent Melissa
A custom agent on Alibaba Qwen 3.5 122B, running entirely on owned hardware: two RTX 6000 Pro 96GB GPUs. No per-token bill, capex plus electricity only, and nothing leaves the building. She is rule six, data stays home, made flesh: private local inference with high context and long-running loops.
Message Melissa on WhatsApp Demo guide and starter prompts →
Questions and answers.
Caoimhe, Melissa and Nara stay live for one week after the workshop, so keep testing them. The deck, the full brief and every source live here at aiagent.research.my. Enquiries: info@basicinsight.my
Download the A4 brief →