MyJackfruit

Help
Browse topics · Building the first step

Building the first step

← Building the first step

How does Kai-Zen run a whole delivery cycle?

Short answer

Kai-Zen picks a step, has it built and checked, puts it live with your go, asks you to check it, then has its Values measured.

Deeper

Also asked:

Do agents hand work to each other? How? Yes. Kai-Zen hands each part to the agent for it: building to the builder, checking to Agent.Scaling, testing to Agent.Tester, vault work to Jackfruit. When one hands its work back, Kai-Zen hands on the next at once; Jackfruit asks Agent.Concept for a missing term.

Kai-Zen has started the first step. What happens now? You mostly wait and answer. Agent.Coder builds the step on your computer, another agent checks it and a separate agent tests it. You are asked when Kai-Zen stops for you, such as before the step goes live.

A delivery cycle, which the other answers call a round of delivery, is how Kai-Zen delivers one step of your project's plan and learns from it. A step, a Value Delivery Step, is the Solutions, or parts of them, delivered together in one short cycle to move some Values. A Value is a perceived benefit, measured as a number, and a Solution is how a Function (what a system does) gets done. An agent is your AI following one written procedure, its standing instructions, for one kind of work.

For a new project, Kai-Zen starts the first round with the first suggested step after it has shown you the plan for one approval, unless you said in your own words that you need not approve it. For a project you already have, a request to deliver runs a round; the answer on picking the next step shows how to ask. A round needs the project's specs, the requirements written in its plan, with each Value's Status (where it stands now), Tolerable (the survival limit) and Goal (the number you plan to reach), and it needs the project's kind and where it is built.

Once the first step has started, you mostly wait and answer. For an app on GilbPlatform, Agent.Coder, the agent that writes your app's code, builds it on your computer; then another agent checks it, and a separate agent tests it. You are asked when Kai-Zen stops for you, such as before the step goes live, and then to check it live. Your AI tells you in the place where it works, GilbApp's chat or the terminal, the window where you type commands; GilbPlatform sends no e-mail or notice. The answers How do I know what my AI is working on right now, and whether it is stuck? and Will I be told when something is ready for me to look at? tell more.

How you know a step is done: in GilbApp's plan, the step's Status shows Building while it is built, Developed only after the tester's saved PASS, and Production once it is proved live. A step counts as delivered only when the people it is for can use it. The round is over when Kai-Zen shows you the measured Values and the next step, as below.

A request to deliver the next step runs the whole round: Kai-Zen picks and writes the step, hands it to the builder, has what was built checked and tested by agents other than the one that built it, puts it live, which it always stops and asks you about first, asks you to check it live, and has each Value the step aims at measured. After each round it shows you the measured Values and the next step. If the spend so far is over Budget, the most of a Development Resource you plan to use, or a Value is still below Tolerable where the plan expected it above, it stops and asks you whether to go on. Otherwise it goes on to the next step, unless you say to re-plan or that you are done. A Development Resource is what the work uses up, such as money or time.

Kai-Zen picks the step for real at the start of the round: it starts from the suggested Value Tracking Table and changes it with what is true today; the answer on picking the next step shows how it weighs the steps. A Value Tracking Table (VTT) follows your Values and Development Resources across your steps: one column for each step, and in each cell the Estimate made before the step is built, then the Actual measured after it is delivered.

It hands the step to the builder for the project's kind: Agent.Coder for an app on GilbPlatform, or the place the project names for anything else. With the step it sends the project's constraints, the things that must be done because a named person requires them, and it first adds the Reuse constraint if the project does not have it yet; the answer on reusing a Solution or a Constraint tells more. If the builder is not ready, Kai-Zen says so, does the steps that need no builder first, and names the date by which the builder is needed. If the builder reports a gap in a spec, the answer on what happens when a spec has a gap shows what Kai-Zen does. GilbApp's chat does not reach the files on your computer unless you choose a code folder while Claude Code is connected; otherwise code kept there needs Claude Code in Terminal. The answer How do I let my AI build in a folder on my computer from GilbApp's chat? shows how.

Then Kai-Zen has what was built checked by Agent.Scaling, the agent that checks what was built, and tested by the builder's tester, Agent.Tester for an app on GilbPlatform, not the agent that built it, with a saved PASS or FAIL. If the builder has no tester, Kai-Zen says so.

Then it puts the step live, into production. It always stops and asks you before a release; when the change alters stored data or how that data is laid out, it first tells you what changes, and only a person, not an agent, can give the go. Without your go it puts nothing live, says what is not live, and goes on with the rest. The answer on how Kai-Zen puts a change live tells more. Once the step is live, Kai-Zen asks you to check it live; it fixes what you send back, puts it live again and asks again, until you approve.

After your approval, Kai-Zen has each Value the step aims at measured with its Meter (how a person, an AI or a test checks it), by an agent other than the builder. A number that was not measured is not a Status, and a reading taken elsewhere than where the stakeholders use it, such as on a local copy, is written beside the Status with its condition, and does not move it; the answer on what counts as proof that something works tells more. A stakeholder is a person, group or system that cares whether a Value goes up or down. Kai-Zen writes the Actuals in the step's column and then decides: keep the Solution, throw it away, or change it. Throwing it away puts each Value's Status back, and the Development Resources used stay used; the answer on deciding to keep or throw away a step's result tells more.

Then it updates each Status (the answer on updating a Value's status shows the same by hand), and sharpens the plan with what was learned: what worked, what did not, how to deliver the Values, and what the stakeholders actually want. The answer on challenging a Solution shows what sharpening does.

In all of this, Kai-Zen hands the work to the other agents in its group, and all vault work to Jackfruit, the vault agent: it finds, reads, writes and links records in your vaults. It gives the agents that each write one level of a plan (Business, Stakeholder or Product; the answer on the levels of a plan explains them) your own words whole, and each of them decides what belongs at its level. When an agent hands its work back, Kai-Zen hands on the next work at once. It reports to you when the work you asked for ends, or when a part stops for one of the acts it always asks you about; a report while work remains is for your information only, and Kai-Zen goes on without waiting for an answer. Each report says what is done, what is not, and why; the answer on reading what an agent hands back tells more.

While work remains, Kai-Zen does not end its turn, and it never asks you to send a message so that it can go on. With Claude Code, a long job can run out of tool rounds, the steps it may take in one turn. The chat then goes on by itself, up to 15 times, and says "The turn ran out of tool rounds, so I am going on by myself (1 of 15). Press Stop to stop." Stop stops it after the turn that is running; send continue to go on.

The answer on answering Kai-Zen's questions lists every act Kai-Zen always stops for, and the one approval it asks for at the end of the work. GilbApp has no control that runs a round of delivery: a round runs when you ask Kai-Zen in the chat.

From GilbApp's chat Kai-Zen works only with the vault, unless you choose a code folder while Claude Code is connected; otherwise a project kept in a folder on your computer needs Claude Code in Terminal. GilbApp keeps your chat with Kai-Zen in this browser tab only, not in your vault. What lasts is what Kai-Zen saves where the project is kept.

This answer is written for a computer screen.

Source: Standards, from the, all, 2026-11-10: Kai-Zen (K kai-zen-06gd2d286dz8s9b6ey63f9e49r) steps Delegate ("Delegate all work to the other agents in this group. Hand all vault work to Jackfruit"), NoPlacing ("Give the level agents the requester's words whole ... Each level agent decides what belongs at its level"), Report ("Take the work back and delegate the next work at once. Report to the requester when the run ends, or when Kai-Zen.AskRules.Stop stops a part. A report while work remains is for information only; do not wait for an answer"); Kai-Zen.Cycle (K 06gdk6x6dhse1fzx2n3cbs91ag) Entry.Specs, Entry.Builder, steps Step, Build, Check, Live, Human, Measure, Update, Improve, Next, Exit.Done; Kai-Zen.New (K 06gdk6xd1hrbdbbyhr6gbtd8wc) Show (unless the requester waived it in their own words), Cycle (After the approval, start the first delivery cycle with the first suggested step); Kai-Zen.Existing.Deliver; Kai-Zen.GoLive (K 06ge4pmrxswnh6nbsh8tcaa3yg) Ask ("If it changes stored data or its shape, tell the requester what changes and ask the requester (a person, not an agent) for go. Without go, stop and say what is not live"), Deploy; Kai-Zen.HumanCheck (K 06ge4t73e1wg32nty0p6eypr08) Ask, Fix; Kai-Zen.AskRules (K 06gdm7dyd5syda52jap5tm8qmm) Stop ("a release (Kai-Zen.GoLive.Ask) ... always stop the part they touch, whatever the requester said ... Stop only that part and go on with the rest"), NoContinue, EndApproval; Kai-Zen.VTTRules.Act (K 06gdzfpv15srd5f0d78btny4b4); Kai-Zen.EvidenceRules (K 06ge5vp5qdx7552qp6hzxe4gq0) Measured, Condition; Kai-Zen.StopRules.Owed; Kai-Zen.ProjectRules.Kind. GilbApp af0f04ee838a4d55887d2ff1f72edd49a4ff7ca0 (https://gilb.com/app/version.json read 2026-11-10, buildTime 2026-11-09.522Z, dirty false), code read: components/agent/AgentChat.tsx:69-70 (AUTO_CONTINUE_MAX = 15), :595-596 (only Claude Code sets budgetSpent), :660-665, :909 ("The turn ran out of tool rounds, so I am going on by myself (n of 15). Press Stop to stop."), :924 (Stop), :931-932 ('Send continue to go on.'). No GilbApp control runs a round: the code shows only Mark delivered (PlanTable.tsx:550, a status mark on a VTT column) and the theme switch. Also: Agent.Scaling, Kai-Zen.New.Down. Kai-Zen.Cycle (Build: the step set to Building; Check: Developed only after the tester's saved PASS; Live: Production after proof at /version or version.json; Next), Kai-Zen.ValueDeliveryStepRules.Delivered, Kai-Zen.UpdateStatus (Planned, Building, Developed, Production, Retired); GilbApp 49b1ae53 (https://gilb.com/app/version.json, read 13 November 2026): src/lib/base/planCard.ts:31 (the five Status words). Live answers, sentences shared: Will I be told when something is ready for me to look at?; How do I know what my AI is working on right now, and whether it is stuck?; When is my app, website or program finished? Code-folder exception: GilbApp 49b1ae53 (https://gilb.com/app/version.json, read 13 November 2026), src/components/agent/AgentChat.tsx:677-692 and :1084-1090 (Choose a code folder, Code in, Claude Code only); live answer How do I let my AI build in a folder on my computer from GilbApp's chat? Live answer How do I have a concept found or made in GilbVeda? (Jackfruit asks Agent.Concept).

Checked against GilbApp d529620.