Chinese programmer discusses Kimi K3
The real friction in AI engineering isn't picking between Loop or Graph Engineering architectures; it's whether you can actually verify the output before execution. A Chinese programmer dissected this by pointing out that sixteen Claude agents managed to build a functional C compiler, though the cost was nearly 2,000 sessions and roughly 2 billion tokens. The core realization is that building the execution loop isn't the hard part. The harness is. Container, Git, Lock, Tests, Recover. that's what makes autonomy survivable. You need a defined specification beforehand because an agent only functions if correctness has been explicitly defined first. Outcomes must be judgeable as pass, fail, or delta to prevent the system from drifting into nonsense. This fundamental issue about verification matters way more than the architectural buzzwords themselves. If you want the full breakdown on how Kimi K3 handles this transition, check out the guide titled 'From Loops to Graphs' for the complete context.
