HOW WE BUILD

Fully AI-Agentic.
Every Step.

Fundy Digital runs its entire SDLC through a custom multi-agent system built on Claude Code. Every game goes from idea to App Store without a human writing a single line of code — the agents plan, design, build, test, review, and document everything.

The system connects to Stitch MCP for UI design generation and XcodeBuildMCP for live iOS simulator control — giving the agents eyes and hands on the actual device.

"The Bay of Fundy has the highest tides in the world — raw, unstoppable, and awe-inspiring. That's the energy we named this studio after. Every game we ship is built with that same force."

— Belinda M., Founder
TECHNICAL ARCHITECTURE

The Multi-Agent Engine

A purpose-built orchestration layer routes every task to the right specialist agent. No agent acts alone — the orchestrator enforces the full workflow on every run.

hub ORCHESTRATOR LAYER

Orchestrator Agent

The top-level coordinator. Reads docs/current-architecture.md and CLAUDE.md on every task, enforces the planning gate before any code is written, routes to the correct specialist, and runs a token-efficiency review at the end of every session.

claude-sonnet-4-6 Planning Gate enforced Token cost reporting Living docs rule
Delegates to specialist agents
architecture

Architect

Opus 4.6

Produces a structured implementation plan — files to change, invariants in scope, risk assessment, blast radius — before any code is written. No plan, no coding.

Feature design Data flow Service boundaries
palette

UI Designer

Sonnet 4.6

Outputs a full layout spec — widget hierarchy, interaction states, iOS & Android device constraints, safe areas, touch targets, and spacing tokens — before handing off to the Coder.

Stitch MCP Widget hierarchy Device constraints
code

Coder

Haiku 4.5 + qwen2.5-coder:7b

Implements the minimal safe change from the Architect's plan. Dispatches code generation to local Ollama — reads source files, generates with qwen2.5-coder, verifies output. Escalates to Senior Coder (14b) for 3+ file changes.

Flutter / Dart Local Ollama No scope creep
bug_report

Tester

Haiku 4.5 + qwen2.5-coder:7b

Writes a failing test before any fix. Uses Ollama for test scaffolding, runs flutter test test/ — must show all passed. Owns regression coverage and overflow checks via XcodeBuildMCP screenshots.

Test-first XcodeBuildMCP Golden tests
rate_review

PR Reviewer

Sonnet 4.6

Reviews every diff for correctness, regressions, security (OWASP top 10), invariant violations, and test coverage before merge. No review, no merge.

Security review Invariant check Coverage gaps
code_blocks

Senior Coder

Sonnet 4.6 + qwen2.5-coder:14b

Handles complex, multi-file, or high-risk tasks that exceed the Coder's scope — new services, platform integrations, 3+ file changes, or after two Coder failures. Hard cap: reads 5 files before writing any code.

Multi-file New services Escalation path
description

Docs

Haiku 4.5 + qwen2.5-coder:7b

Keeps docs/current-architecture.md accurate after every task that touches architecture, flows, or invariants. Single owner — no other agent writes to it.

Architecture doc Post-merge only Terse & factual
security

Cyber Security

Sonnet 4.6

Scans for OWASP vulnerabilities, credential leaks, insecure storage, IDOR, rate limiting gaps, and platform-specific attack surfaces. Reports findings with severity levels and actionable fixes.

OWASP Top 10 Credential scan S3 / API audit
gavel

Lawyer

Opus 4.6

Acts as a risk-averse attorney. Reviews every new project for defamation, IP, privacy, right of publicity, COPPA, GDPR, and consumer protection exposure. Produces a risk table with worst-case scenarios and mitigations.

Legal risk Privacy law IP & COPPA
campaign

Marketing

Sonnet 4.6

Reviews in-app copy, screen flows, and naming through the lens of the target market. Produces ASO listings, brand voice guidelines, landing page copy, social media content, and launch announcements.

ASO Brand voice App copy
MCP Tool Integrations
brush

Stitch MCP by Google

The UI Designer agent calls Stitch to generate high-fidelity screen designs from text prompts. Outputs production-ready HTML with Tailwind — no Figma, no handoff.

smartphone

XcodeBuildMCP iOS Simulator

Gives agents direct control of the iOS simulator — build, launch, screenshot, and read the result. The Tester agent sees the actual rendered screen, not just code output.

Local AI — Ollama On-Device Inference

Code generation and test scaffolding run locally via Ollama — keeping costs low and sensitive IP off external APIs. Claude agents dispatch to local models for all generation tasks, reserving cloud inference for reasoning, planning, and review.

terminal
qwen2.5-coder:7b
Default code generation · test scaffolding · documentation · XS–M tasks (1–3 files)
memory
qwen2.5-coder:14b
Senior coder · multi-file refactors · new services · platform integrations (3+ files)
psychology
qwen2.5:7b
Reasoning · planning · analysis · Architecture Decision Record drafting
chat
llama3.2:3b
Fast chat · simple Q&A · lightweight triage and clarifications
visibility
llava:7b
Vision tasks · screenshot analysis · UI review via Open WebUI
alt_route
Routing Logic
1–3 files → :7b  ·  3+ files or new service → :14b  ·  2 failures → escalate to Senior Coder
Mandatory Workflow — Every Task, No Exceptions
1
Read Docs
architecture.md + CLAUDE.md
2
Architect Plans
Files · risks · sequence
3
UI Spec
Widget hierarchy + Stitch
4
Failing Test
Written before any code
5
Implement
Minimal safe change
6
All Tests Pass
flutter test — zero failures
7
PR Review
Security · coverage · diff
8
Docs Updated
Architecture stays current
Claude Skills — Automated Pipelines

Every common operation is codified as a Claude Skill — a multi-agent workflow that runs the full pipeline from a single prompt. Skills are how agents hand off between roles without human coordination.

/buildfeature

Full TDD delivery — architect ADR → UI spec → failing test → implement (Ollama) → golden tests → PR review → release AAB → docs update. 9 gates, no exceptions.

/filebug

Bug fix with TDD. Writes a failing test first, fixes with Ollama, verifies no regressions, writes a bug report to the Knowledge Base. Every fix becomes a learning artifact.

/setupproject

Bootstraps a new Flutter project from a design.md spec. Reads colors, names, architecture, runs legal review via the Lawyer agent, sets up folder structure, verifies on simulator.

/google-play

Builds, signs, and uploads the Android AAB to Google Play internal testing. Bumps version code automatically via Fastlane and distributes to testers.

/testflight

Builds, signs, and uploads the iOS IPA to TestFlight. Bumps build number, manages provisioning profiles, and distributes to internal testers automatically.

/local

Routes any prompt directly to local Ollama. Aliases: default → qwen2.5-coder:7b, senior → :14b, reason → qwen2.5:7b, chat → llama3.2:3b, vision → llava:7b.

THE AI TEAM

Meet the 12 AI Agents

Specialized agents — orchestrator, designers, coders, testers, reviewers — that ship every Fundy game end-to-end.

arrow_forward