A Malaysian hosting company is putting an AI agent on every desk

A conversation with someone at a hosting company gave me a look at what Malaysian companies are actually experimenting with in the AI agent space.

The plan: an agent for every employee, everyone gets their own harness, skills shared across the company. Everything self-hosted because they don’t want the codebase touching a public model. Serious GPU hardware, and more arriving later.

The model story is the part I keep thinking about. They run a self-hosted mixture-of-experts model, and the team consensus is it’s 有点笨, so they’re testing alternatives. When I asked how they evaluate models, the answer was basically “we ask the dev team how it feels.” No benchmarks, no eval sets. A hosting company with serious hardware, picking models by vibes.

Meanwhile my company is doing the opposite. No GPUs at all, just a subscription and an API. We’re experimenting with Hermes profiles so each customer company gets its own sandbox: a separate bot, separate memory, nothing shared between customers. One investigator bot per customer instead of one per employee. And the model decision went the other way too: my CEO is testing the tools himself. He’s impressed with Codex Luna, so that’s the direction we’re moving.

That’s the part that sticks with me. One company has the hardware and evaluates by asking around. The other has no hardware and evaluates by having the boss actually use the thing. The guy with the GPUs is still guessing. The guy with a subscription just found something that works.

Everyone is building MCP tools. Nobody is benchmarking. But maybe the benchmark is simpler than we think: hand it to the person who pays for it, and see if they’re still impressed after a week.