Can I trust what the AI suggests and builds?
Short answer
Check it: Kai-Zen shows the whole plan for your approval, has each piece tested separately, and puts nothing live without your yes.
Before you start
Words used here (1)
- Kai-Zen
- Kai-Zen is the planning agent: it finds out what you want, chooses the way, hands the work to other agents and reports back.
Deeper
An AI can be wrong, so do not take what it says on trust. The way Kai-Zen works is built so that you can check it.
What it suggests: Kai-Zen asks what only you know at the start, and decides the rest itself, such as the way of working, the tools and the estimates; you may change them later. What only you could have told it but it had to choose, it writes down as an assumption, a choice it made for you that you can change. Before it builds any screen, it shows you the whole plan for one approval, unless you said in your own words that you need not approve it: the steps it suggests with what each should bring and cost, every assumption, and every open point. Reply with what you want changed; it changes it and asks once more.
What it builds: Agent.Coder, the agent that builds apps, has each piece tested by Agent.Tester, the agent that tests what others built, started as a separate agent that knows nothing of how it was built. If it has to test a piece itself, it says so. Agent.Scaling, the agent that checks what was built, checks it too. Each result is marked REAL if it ran on the real thing, MOCK if it ran on a stand-in, or UNPROVEN if nothing was run; the answers on what counts as proof that something works and on reading what an agent hands back explain them. Then Kai-Zen asks you to check each change live yourself, and fixes what you send back until you approve.
Can it build something you did not ask for? Kai-Zen has the approved plan built, one step at a time. For a change you name, it has only that change built; anything else it sees, it notes and does not do. A few things come with every app on GilbPlatform without your asking: the first step sets up a ready-made core app with a sign-in, and Agent.Coder makes the pictures the app needs, such as a logo, itself.
Some acts it never does without you: putting work live, ending a record or deleting a vault, giving access to a vault, spending more than the budget, and changing a goal you set. Agent.Coder never puts an app on the internet without your yes. The answer on why Kai-Zen stops and asks lists them all.
When your AI asks you to say yes. Two kinds of question come. Kai-Zen asks for your yes on your project, such as the plan or putting a change live: read what it says will happen, say yes only if that is what you want, and otherwise say what to change. Your AI coder, the AI that writes your app's code, may ask to allow an act on your computer. In a terminal, the window where you type commands, Claude Code or Codex may ask to allow a tool or a command before it runs; refuse what you do not want. In GilbApp's chat, with a code folder (the folder on your computer where your AI writes your app's code) and Approve before each write, it asks Allow this AI to run this command? with Allow and Decline. A command that installs, builds or runs your app is normal while building: allow it when it fits what you asked for. Refuse it when you do not understand it or it has nothing to do with your request, then ask your AI what it does. When your AI asks to work in a vault, a place where your records are kept, choose yes for this time, not always: each vault switch stays your decision. Anthropic says: "You're responsible for reviewing proposed code and commands for safety before approval." (https://code.claude.com/docs/en/security, read 13 November 2026).
GilbApp does not check these rules: they are the agents' written procedures, and the agents follow them. If something looks wrong, say so in the chat, or write to [email protected].
This answer is written for a computer screen.
Source: Standards Kai-Zen.AskRules (Decide, MustAsk, Assume, Stop, EndApproval), Kai-Zen.New (Standard, Show), Kai-Zen.ValueDeliveryStepRules.First, Kai-Zen.Cycle (Build, Check, Human), Kai-Zen.Update (Build: Build only this change. Note anything else you see; do not do it, as cited by the live page on changing the product) and Agent.Coder (AfterIncrement, GilbStack 7a images, AskGoLive Never publish without the person's yes), read from GilbPlatform on 12 November 2026. Shared wording: live answers Why does Kai-Zen stop and ask, or follow a standard?, How do I read what an agent hands back?, What counts as proof that something works?, How do I build an app with an AI coder?, How do I change the product while a project runs?. GilbApp 49b1ae53 (https://gilb.com/app/version.json, read 13 November 2026): src/components/agent/CodeFolderParts.tsx:93 (Approve before each write), :126-140 (Allow this AI to run this command?, Allow, Decline). Anthropic, https://code.claude.com/docs/en/security, and OpenAI, https://developers.openai.com/codex/agent-approvals-security, both read 13 November 2026. Live answers, sentences shared: What is my AI doing when it says building or running tests?; Can my AI break or delete my data? How do I stay safe?; How do I make an AI vault and connect Codex to it? (yes for this time, not always).
Checked against GilbApp d529620.