MiniMax integrated into agents; set aside AGI for now

A friend recommended MiniMax to me, so I'm trying it out to see if it's actually good. This week, during a client pilot, I integrated MiniMax into WorkBuddy's agent workflow. The client needs to process Chinese meeting minutes and expense reports. Overseas models can handle it, but the review costs are annoying.

My main focus this time is whether it fits into the workflow; let's not discuss how far AGI is yet. MiniMax has been getting a lot of buzz recently. Reports say founder Yan Junjie won't take a salary until AGI is achieved, and he defines AGI as AI contributing 1% of global GDP. AGI, or Artificial General Intelligence, sounds huge. That's also the problem: the AGI narrative and engineering usability aren't the same thing. In projects, we only care if tasks run stably, like converting meeting notes into to-do items.

I followed the standard path. Went to the console, created an API Key, selected a chat model, and connected it to WorkBuddy. An API Key is essentially a temporary key for calling the model. I also used MCP to connect local document permissions; MCP allows the model to read external tools. I've used it for four weeks before, and now I'm testing tool calling with it. I wrote strict prompts: first extract names, dates, amounts, and to-dos, then output JSON. JSON is structured text readable by machines, allowing downstream systems to automatically ingest the data.

Here are some impressions from running it. The surprise was in Chinese expression. For colloquial phrases in meeting notes like "Old Wang said he'll submit that expense report next week," MiniMax can organize them into to-dos without stiff translation, and the tone feels natural. The bottleneck is in the engineering pipeline. The entry points for console documentation are scattered, and I'm unsure about default model parameters. Compared to Claude and ChatGPT's plugin ecosystems, MiniMax's toolchain is still thin. During tool calls, it occasionally misses fields, requiring extra validation. Audit logs aren't as clear as mature SaaS platforms, which isn't reassuring for compliance. Regarding cost, the official team keeps emphasizing inference cost optimization. My tests show calls aren't slow, but I haven't gotten precise enough numbers to put into a quote sheet.

The conclusion depends on the situation. It's suitable for small teams doing Chinese dialogue, meeting summaries, lightweight copywriting, and who are willing to do their own validation. It's not suitable for processes with strong compliance, strict permissions, or direct invoicing/expense reimbursement integration. Especially don't push it to production just because of the word "AGI" in the news.

Next steps are simple: find a dozen real tasks, skip the official demos, run a round, and keep the failure samples. Scale up if it works. You can listen to the model narratives, but you need to look closer at bills and audits.