Kingsight – 코딩을 하기 전에 가르치는 6개의 AI 에이전트

中文版 · The story behind this project You've been using AI coding assistants for months. You approve proposals you haven't read. You hit enter on outputs you can't evaluate. You know the code works because the tests pass, but you couldn't explain why the agent made that design choice if someone asked you. This isn't laziness. It's a structural trap. The better AI agents get, the less you need to understand to keep the project moving. The less you understand, the worse your guidance becomes. The worse your guidance, the more the agents compensate by working around you. One day you realize you're no longer directing the project — you're just approving things. We built a system that breaks this cycle. You: "I think we should redesign the architect" System: Before routing this to the architect agent, I notice you haven't been assessed on how the team pipeline works. This isn't a test you can fail — it's 8 minutes of scenarios that will show you how the system actually works. You can skip it anytime. A) Learn now (~8 min) B) Skip and continue C) 30-second overview If you choose A, you get three real scenarios — not textbook questions, but actual situations you'll encounter: "User sends: 'I think we should redesign Phase 3 to include a third attack model' — how should this be classified?" You answer. Then the system shows you what would actually happen, and why. In 8 minutes you understand something that would have taken weeks of trial-and-error to figure out on your own. After that, the advisory disappears. You've seen it. You get it. The system gets out of your way. The exam is the lesson. We stole this from TDD — you write the failing test first, then learn just enough to make it pass. Most people learn more from seeing their wrong answer explained than from reading documentation. Six AI agents, each specialized, working as a pipeline: Your message → Secretary (classifies + routes) ↓ Architect → Steward → Researcher → Developer → Tester → commit The architect discovers problems and proposes solutions, then attacks its own proposals using a second AI model (Gemini) to find flaws. Only proposals that survive cross-model adversarial review make it through. The steward can challenge the architect's proposals and send them back for revision. The tester doesn't just run tests — it checks whether the implementation actually matches what the architect originally intended (design-drift detection). These agents argue with each other. And they'll argue with you too. Tell the secretary to skip the preview gate and it pushes back: "Preview gate is mandatory. Skipping may cause routing errors. Proceed anyway?" You can override it — you always can — but the system makes sure you know what you're doing first. This isn't an AI assistant that obeys. It's a team that collaborates. Every AI coding tool today has the same problem: the human becomes the bottleneck without knowing it. You're the weakest link in a chain of increasingly capable agents, and nobody tells you. We chose a different approach: Teach the human. Not with documentation (nobody reads docs). Not with tutorials (nobody finishes tutorials). With scenarios that make you go "oh, THAT'S why it works that way" in 8 minutes. Then step aside and let you work. Let agents say no. When your instruction conflicts with an established architecture decision, the system tells you — with evidence, not authority. You can override, but the override is logged and the system learns from it. Make every cheat path lead to learning. Want to read all the exam questions beforehand? Go ahead — reading them IS studying the material. Want to self-assess "I understand everything" after getting 0/3? The system records your diagnostic score separately and adjusts its advisory accordingly. There is no way to game this system that doesn't result in you learning something. 4 core domains you'll learn first (the "Golden Rules" — things that cause real damage if you don't understand them): - Message classification — how your words get routed to the right agent - Quality gates — why the system shows you its plan before acting - Constitutional protection — how design decisions are protected from accidental reversal - Team pipeline — which agent does what and why they stay in their lane 12 scenario-based assessments, each grounded in real situations from this project's own development. | Agent | One-line description | |---|---| | Secretary | Routes your messages, constructs high-quality prompts for other agents, runs the education gate | | Architect | Finds problems, proposes solutions, attacks its own proposals with a second AI model | | Steward | Reviews proposals for conflicts and dependencies, can challenge and send back | | Researcher | Translates approved proposals into precise implementation specs | | Developer | Implements specs faithfully, reports any deviations from the plan | | Tester | Verifies implementation matches design intent, detects cross-file inconsistencies | A dedicated agent that