This week we noticed a signal: Chinese AI coding tools are collectively catching up on "planning capability." Moonshot AI's Kimi Code officially launched Plan Mode in its latest update. The core change can be summarized in one sentence — when AI receives a complex task, it must first submit an execution plan, and only act after the user approves it.
What this is
Plan Mode addresses a simple but high-frequency problem: AI moves too fast. Ask it to refactor a module, and it immediately starts editing files. Halfway through, it realizes it misunderstood your intent, and everything has to be rolled back.
The new Kimi Code introduces an intermediate step: when faced with a complex task, AI first enters a "read-only" state, analyzes the code, breaks down the steps, flags risks, and generates a structured execution plan — including objectives, steps, expected output per step, and verification methods. Users can approve, modify, or reject the plan. Only after approval does AI actually act, and every step has a checkpoint.
From an engineering perspective, this essentially turns "Human-in-the-Loop" from a slogan into a procedural constraint. The plan is not disposable text once written — it's a stateful object that can be tracked across conversation turns, resumed from breakpoints, and audited afterward.
Industry view
Supporters see this as a critical step toward Agent maturity. The dominant voice in the developer community: the models themselves are already smart enough; the bottleneck is "preventing AI from running wild." Plan Mode is like adding brakes and a steering wheel to AI — only with these can it be safely deployed in enterprise production environments.
But there are sober counterarguments. One view holds that Plan Mode looks rigorous but actually shifts the planning cost to the user — every time requirements change, the user must review the plan again, which becomes a burden for complex tasks. A sharper critique, drawn from Anthropic and OpenAI's early practices: over-structured plans make AI rigid, unable to adapt to legitimate branches outside the plan, ultimately performing worse than letting the model set its own pace. In other words, human-in-the-loop is sometimes a safety net, sometimes a straitjacket.
Our judgment: the direction is right, but Plan Mode is not the endpoint. The next competitive frontier is "how good the plan is," not "whether there is a plan."
Impact on regular people
For enterprise IT: If you're evaluating private deployment of AI coding tools, Plan Mode is a must-ask item. An Agent without a plan review mechanism is effectively running naked on production codebases.
For individual professionals: People using AI for automation and scripting will feel the difference first — instead of repeatedly prompting to correct direction, you let AI propose a plan and confirm it upfront. The efficiency gain is real.
For the consumer market: No visible change in the short term, but this is a signal that Chinese AI tools are moving from "runs" to "production-ready," which will accelerate enterprise procurement decisions.