An Operator’s Year with AI Agents: How Tso Chinese Built Its AI Stack in 2026
Angell Tsang, Chief AI Officer and Co-Founder at Tso Chinese, looks back on a year of building with AI agents. He covers automating catering sales, sending 5-star reviews from the Marqii MCP to product pages, and having his agents write the bug reports to his vendors.
It was very tempting for me to start this post by asking my AI agent to write it for me, but I decided I had to write this myself. Now, that being said, I'm not actually writing anything. I'm just speaking to my Wispr Flow app, and it's writing this post for me. For the record, this entire post is a transcription of spoken word.
2026 is going to go down as a wild year. If you had told the November 2025 version of me what I’d be doing with AI, the Marqii MCP, and other vendors' MCPs/APIs, I wouldn't believe you. The growth, change, and improvement possible for anyone armed with today’s APIs has been remarkable.
My personal journey with AI started in January 2026 with OpenClaw, which truly blue-pilled me into integrating my work with AI agents.
The first thing I did was write some simple automation to handle some of our basic catering sales functions. After that went well, I thought to myself, "Well, what else do I want to fix, add, or create?"
I wasn't about to go and touch my actual main codebase quite yet. This was in the spring of 2026, and OpenClaw was still pretty new. It was still fairly unstable, but I got an early taste of the dopamine hit that you get when you prompt something into existence.
The second little project I ended up building was taking customer reviews from the Marqii MCP connection, adding them to Airtable as a quasi database, then building a Slack app to connect it together, and then connecting it all to my frontend website to display 5-star reviews for products. So then, whenever someone leaves a 5-star review for a certain product, this automation would take a look at the product mentioned in that review, clip it, and shoot a message into my Slack app asking me if this passed my approval or not. I then could leave my comments on why I did or didn’t like that particular review, and if approved, that snippet would get posted onto the product detail page with the reviewer's first name and a timestamp for authenticity.
I (as in, my agent) got it working, but it was still fairly rudimentary code. By May, I had decided to migrate from the OpenClaw harness to Hermes, which was one of the best decisions we made. I ended up adding a Pi harness for most coding tasks, and Hermes would delegate coding tasks to Pi, and by summer, that code was polished up.
Then, in September of this year, we launched a full rewrite of our website, https://tsochinese.com. I was then able to use our agents to rewrite the entire Marqii reviews automation for the new website.
There were some early issues with the Marqii MCP (as expected with new releases), and we sent feedback to the Marqii team through my agent. Of course, I didn't write a single email! Every email was sent to the Marqii team from our agents, and the Marqii team replied to my agents with comments and fixes to their MCP (I’m sure they knew this was happening). My agents read those emails and immediately turned around and started working on implementing their changes and fixes.
That's pretty much how I’m interacting with the technical teams of all of my vendors now. I'm not really writing emails anymore. My agents are writing those emails specifically around feature requests and reporting bugs. Some of my prompts to my agents look like this: "Hey, if anything is missing in this vendor's MCP, what recommendations would you give them? What could make their MCP even better to give us better compatibility, stability, and reliability?"
This is the future of work. It's never been a more exciting time to build anything you can imagine.
In January, I had three repositories in my GitHub account, and now we have 67. Each one represents a problem solved, some tiny and some massive, both in scope and in what it would have cost to build the traditional way.
If you haven't yet done anything like this for your business, it's easier than you think to get started. Just tackle one small problem at a time. Get a small taste of what it's like to solve problems quickly and easily, and you're only going to want to build more and bigger.
Angell Tsang is the Chief AI Officer and co-founder at Tso Chinese Takeout and Delivery (tsochinese.com), based out of Austin, Texas. He leads the technology innovation efforts at his five restaurants, all running on a fully custom software platform.
Not working with Marqii yet? Schedule a demo to see how Marqii helps your restaurant get discovered and get chosen.
More resources:
Introducing Marqii MCP and How to Set It Up.
10 Things You Can Do in Marqii Without Leaving Your AI Chat
How Restaurants Are Using The Marqii MCP To Work Smarter
Why We Built the Marqii MCP: Giving Operators Back Control of Their Data