Caffeine V3 is an agentic build system. You no longer talk to a single model — you direct a team of four specialized agents that work in parallel to turn a one-line idea into a deployed app.
This section explains how the agent team collaborates, how the build loop works, and why context is managed like RAM in V3.
> THE AGENT TEAM
Four specialized agents each own one phase of the build. They share the same context, so one agent's output becomes the next agent's input.
Composer
Planning
Breaks a one-line idea into a structured build plan — features, data model, and architecture.
Frontend
Interface
Builds the React interface from the plan — pages, components, and design system.
Backend
Canister
Writes the Motoko canister logic — state, queries, and update methods.
Quality
Verification
Compiles, type-checks, and verifies the build — catching errors before deployment.
> HOW THE BUILD LOOP WORKS
Every build follows the same loop. The agents do not generate everything at once — they iterate, verify, and self-correct.
Plan
The Composer reads your prompt and produces a structured build plan.
Build in parallel
Frontend and Backend agents work at the same time, each on its own layer.
Verify and retry
The Quality agent compiles and checks the build. On error, the loop retries with a fix.
Iterate
Each round brings the app closer to a runnable state until the build passes.
Deploy
Once the build passes, the app is deployed to ICP — no manual steps.
> CONTEXT MANAGED LIKE RAM
In V3, context is a limited resource. The agent team keeps only the files relevant to the current task in active context — the way an operating system keeps hot data in RAM.
This means large projects do not slow down from context overflow. The agents load files on demand, do the work, then release them.
[KEEP THE PROMPT FOCUSED]
> PREVIEW TOOLS
V3 gives you two ways to review your app while it builds — a side-by-side preview pane and preview comments with snapshots.
[SIDE-BY-SIDE PREVIEW PANE ON DESKTOP]
[PREVIEW COMMENTS WITH SNAPSHOTS]
> WHAT CHANGED FROM V2
From a single model to a team of four specialized agents.
Frontend and Backend now build in parallel instead of sequentially.
Context is managed like RAM, so large projects stay fast.
The Quality agent compiles and verifies every build before deployment.
The build loop self-corrects — errors retry internally instead of failing outright.
> COPYABLE PROMPTS
These prompts show how to write for the agent team — clear, structured, one feature at a time.
Build a habit tracker for personal use.
Core features:
- Add a habit with a name and a daily target
- Mark a habit complete for today
- Show a streak count per habit
Data model:
- Habit: id, name, target, createdAt
- CheckIn: habitId, date
Auth: Internet Identity
Architecture: single canisterAdd a weekly summary view to the habit tracker.
- Show completion rate per habit for the last 7 days.
- Add a bar chart using Recharts.
- Do not change the existing habit list or check-in logic.Change the habit list so each row shows the streak count on the right.
Currently: the streak is only visible on the detail view.
Should be: a compact streak badge on every row in the list.
Do not change the check-in button or the data model.> V3 BUILD CHECKLIST
- I understand V3 uses four specialized agents instead of one model
- I know Frontend and Backend build in parallel
- I understand context is managed like RAM
- I know the Quality agent verifies every build before deployment
- I keep my prompts focused — one feature at a time
- I know the app deploys to ICP automatically once the build passes
[THINK LIKE A TEAM LEAD]