Chen Kai
Senior agent engineer at Alibaba International · harness and knowledge engineering
I am a senior agent engineer at Alibaba International, responsible for the agent infrastructure behind the ads business. The work runs on four lines. A reporting dashboard turns the weekly reports into a pipeline that can be plugged in, compared, and signed off, with ingest, version diff, daily patrol, review, and release approval chained together and every step leaving evidence. A domain knowledge base is built on the ads business ontology, split into raw material, static knowledge, and application packs, with a glossary and taxonomy to keep terms consistent and a citation gate so answers point back to source text. A question-answering harness keeps improving through a fixed question set, side-by-side runs of the same questions, and traces flowing back from production, so every change can be measured. An internal agent chat platform connects colleagues' local knowledge bases to the agent and links them with the ads QA service, with read-only grants, zero access by default, per-person visibility, and multi-agent handoffs.
All four rest on the same judgment: the model sets the floor, and the harness sets the ceiling. An agent that goes to production needs memory it can verify, evaluations it can compare, deliveries it can roll back, and permissions it can hold. My habit is to cut a business problem down to what a machine can judge, then let the agent finish it.
Before this I was an AI knowledge engineer at Alibaba Cloud, running documentation operations for Model Studio (Bailian) International. Four agents took over most of the daily work in a DingTalk group; the team ended up with 30+ production Skills, about 95% task automation, and 100+ hours saved every week. After hours I build my own systems, mostly because I wanted a tool that did not exist.
What I do now, and what I did before
Agent infrastructure for the ads business: reporting pipeline, domain knowledge base, QA harness, and an internal agent chat platform.
Documentation operations for Model Studio (Bailian) International, run by four agents; 30+ production Skills and about 95% automation.
M.S. in Applied Mathematics and Statistics, Johns Hopkins University (2023.08–2024.12) · B.S. in Mathematics, Sun Yat-sen University (2018.09–2022.06)
The harness ideas were tested in my own projects first
Putting an agent into production is less about the model and more about what surrounds it: a knowledge base that can cite its sources, a fixed set of questions to compare each change against, traces coming back from real traffic, and permissions that hold when someone else's data is involved. I had already built those pieces in FileVault, NodeMind and DaaS before the ads work needed them, so most of the design was a matter of moving them over. Edge cases from work go back into the projects.
See the systems, one by oneCan an agent answer a business question without making things up, and show where the answer came from?
What that takes: knowledge cut to what a machine can judge, every answer traced back to its source, a fixed question set to compare each change against, and permissions that hold across people and agents.
A few principles I code by
- Judgment firstRules exist to be broken, as long as you can say why. No rule outranks the specific call in front of you.
- Write it downIf you cannot explain it in prose, you have not thought it through yet. The design lands as text first; code is one way to implement it.
- No early abstractionThree similar lines beat one guess at the pattern. Wait for the pattern to show up, then close it up.
- Stability over flashA week without an incident matters more than a good demo. Most of the work is in the last mile.
- Agents are systemsThey fail, they need recovery, and someone has to read the logs. Design them like distributed systems, not chat toys.