Chen Kai
Engineer · Writer · Building long-running agent systems @ Alibaba Cloud
I build long-running agent systems that recover from their own failures — the kind that take over an entire workflow and run for days or weeks, not one-question-one-answer chatbots.
Underneath, it’s one set of problems: getting several models to divide work and check each other instead of grading their own homework; compressing each failure into something reusable; keeping shared memory from drifting across sessions; and closing the gap between a demo that works and a system you’d actually leave running overnight.
I also write technical articles, separately in Chinese and English — though that’s a byproduct.
Marketing, research, games — different surfaces, one machine underneath
My work spans surfaces that look unrelated — cross-border marketing, autonomous research, game generation, agent control planes. To me they're one problem wearing different clothes: getting a crowd of models to divide labor and correct each other, compressing every failure into something reusable, and keeping a system running honestly for days or weeks without collapsing. Each new domain is the same foundation, tested once more against reality.
See the systems, one by oneHow do long-running systems stay resilient through failure, model swaps, cost pressure, and infrastructure migration?
Concrete levers: dynamic token budget across providers, compressing failures into reusable Skills, type-converging shared memory, and bridging the gap between demo and production-ready observability, replay, and trust.
A few principles I code by
- Judgment firstRules exist to be broken — once you can say why. No principle outranks the specific call in front of you.
- Docs firstIf you cannot write it down clearly, you have not thought it through. The design lands as prose first; code is one implementation of it.
- No early abstractionThree similar lines beat one premature abstraction. Let the pattern surface on its own, then close it up.
- Stability over flashRunning for a week without collapsing beats a dazzling demo. The hard part is the last mile, not the first screen.
- Agents are systemsDo not treat an agent as a chat toy — design it as a long-running distributed system that fails, recovers, and audits itself.