01 Communication Architecture

Three distinct channels serve different coordination patterns: real-time messaging, async task tracking, and topic-based group conversations.

flowchart TD
    subgraph CH1["sessions_send โ€” Direct Messaging"]
        P1["Prowl ๐Ÿš”"] -->|digest JSON| W1["Wheeljack ๐Ÿ› ๏ธ"]
        O1["Optimus ๐Ÿ "] -->|task delegation| W1
        O1 -->|status query| P1
    end

    subgraph CH2["Fizzy Cards โ€” Task Tracking"]
        FC["Fizzy Board"]
        FP["fizzy-pop
webhook daemon"] FM["fizzy-md
Markdown proxy"] FP -->|real-time events| FC FM -->|HTML rendering| FC end subgraph CH3["The Ark โ€” Telegram Group"] T1["Coding Topic"] T2["Family Topic"] T3["Life & Finance"] T4["Business & Strategy"] T5["General / Commands"] end CH1 -.- CH2 CH2 -.- CH3 classDef ch1 fill:#1e3a5f33,stroke:#2563eb,stroke-width:2px classDef ch2 fill:#04785733,stroke:#059669,stroke-width:2px classDef ch3 fill:#b4530933,stroke:#d97706,stroke-width:2px classDef agent fill:#1e3a5f22,stroke:#60a5fa,stroke-width:1.5px classDef tool fill:#04785722,stroke:#34d399,stroke-width:1px classDef topic fill:#b4530922,stroke:#fbbf24,stroke-width:1px class CH1 ch1 class CH2 ch2 class CH3 ch3 class P1,W1,O1 agent class FC,FP,FM tool class T1,T2,T3,T4,T5 topic
Ark Group Chat โ€” Default Listeners
๐Ÿ› ๏ธ
Coding
โ†’ Wheeljack
๐Ÿ
Family
โ†’ Bumblebee
๐Ÿš”
Life & Finance
โ†’ Prowl
๐ŸŽฏ
Business & Strategy
โ†’ Bluestreak
๐Ÿ 
General / Commands
โ†’ Optimus
๐Ÿ“Š
Reports
โ†’ Fuel checks, usage reports
sessions_send

Direct agent-to-agent messaging. Structured JSON payloads. Used for handoffs where data format matters โ€” like Prowl passing digest JSON to Wheeljack.

Fizzy Cards

Task board with tags: bot-actionable needs-human. fizzy-pop delivers webhook events in real-time. fizzy-md converts Markdown descriptions to HTML.

The Ark Group

Telegram supergroup with forum topics. Each topic has a default listener agent. Agents can also listen cross-topic when mentioned.

02 Knowledge Management

Shared filesystem as the single source of truth. Access is scoped by domain โ€” not everything sees everything.

knowledge/
โ”œโ”€โ”€ shared/ all agents
โ”œโ”€โ”€ personal/ Optimus + Prowl
โ”œโ”€โ”€ work/ Optimus + Wheeljack
โ””โ”€โ”€ family/ Optimus + Bumblebee
"Facts are never deleted, only superseded." โ€” Each entry carries a timestamp. New information overwrites old context without destroying history.
Memory Pipeline
Daily notes
โ†’
MEMORY.md
โ†’
Weekly extract
โ†’
knowledge/
Daily Notes
memory/YYYY-MM-DD.md โ€” raw session logs, every agent writes
MEMORY.md
Curated long-term memory, Optimus maintains
Weekly Extraction
Structured facts promoted to knowledge/ files
๐Ÿ…ฟ๏ธ Park It โ€” Routing Decision Tree

When parking work, the destination depends on what kind of content you're storing. Multiple destinations can combine.

Park It ๐Ÿ…ฟ๏ธ
โ”œโ”€โ”€ Tracked task? โ†’ Fizzy card update steps + desc + column
โ”œโ”€โ”€ Lesson / decision? โ†’ MEMORY.md curated, โ‰ค4KB
โ”œโ”€โ”€ Reference material? โ†’ knowledge/ structured facts
โ”œโ”€โ”€ WIP context? โ†’ memory/YYYY-MM-DD.md raw daily logs
โ””โ”€โ”€ Needs revisiting? โ†’ Cron reminder scheduled wake
๐ŸŽฏ Fizzy Card
Update existing card โ€” check off steps, update description, move column. Never create new cards when parking.
๐Ÿง  MEMORY.md
Only for genuinely reusable insights, convention changes, or important decisions. Strict size budget.
๐Ÿ““ Daily Notes
Default destination for most parking. "Here's where I left off." Archived after ~7 days.
โฐ Cron Reminder
Schedule a future wake event to revisit parked work at the right time.
03 Cost Optimization Journey

Weekly usage relative to 4-week rolling average (= 100%). Bars show how each week compares to our normal operating cost.

Feb 22 โ€” Model tiering + free fallbacks 99%
99%
Mar 1 โ€” Claude Max plan stabilized 52%
52%
Mar 8 โ€” Heavy agent work (new bots onboarding) 250%
250%
Mar 15 โ€” Normalized 76%
76%
Mar 22 โ€” Low activity (Lebaran) 9%
9%
Mar 29 โ€” Dispatcher built, viz pages created 64%
64%
Model Assignment Table
Agent Primary Model Weekly Cost
๐Ÿ  Optimus Opus 4.6 45%
๐Ÿš” Prowl Opus 4.6 19%
๐Ÿ Bumblebee Opus 4.6 12%
๐Ÿ”ฌ Perceptor Opus + Sonnet 7%
๐Ÿ› ๏ธ Wheeljack Codex Free 4%
๐Ÿšจ Red Alert Opus + Sonnet 3%
Others (7) Sonnet / Opus 9%
Free Tool Layering
Amp
Free daily credits
replenish hourly
OpenCode
MiniMax, GLM, Kimi
always free
Codex CLI
Paid fallback
when free exhausted
Deterministic tasks use launchd for zero tokens. Only tasks requiring judgment burn model credits.
04 Agent Interaction Map

Optimus is the central coordinator. Some agents are fully connected; others are deliberately air-gapped.

flowchart TD
    OP["Optimus ๐Ÿ 
Coordinator"] --> PR["Prowl ๐Ÿš”
Life Intel"] OP --> WJ["Wheeljack ๐Ÿ› ๏ธ
Engineering"] OP --> BB["Bumblebee ๐Ÿ
Family"] OP --> BS["Bluestreak ๐ŸŽฏ
Strategy"] OP --> RA["Red Alert ๐Ÿšจ
BTA Night Watch"] OP --> HN["Hound ๐Ÿ”
Tax"] OP --> PC["Perceptor ๐Ÿ”ฌ
Research"] OP --> RW["Rewind ๐Ÿ“ผ
Blog"] OP --> JZ["Jazz ๐ŸŽท
Public Voice"] OP --> RV["Rivet ๐Ÿ”ฉ
Vika's Organizer"] OP --> AT["Alpha Trion ๐Ÿ“ฟ
Islamic Knowledge"] PR -->|digest pipeline| WJ WJ --> AMP["Amp"] WJ --> CDX["Codex CLI"] WJ --> OC["OpenCode"] RT["Ratchet โค๏ธโ€๐Ÿฉน"] -.- AIR1["Air-gapped"] AC["Arcee ๐Ÿ’—"] -.- AIR2["Air-gapped"] FIZZY[("Fizzy
Task Board")] ARK[("The Ark
Group Chat")] OP --- FIZZY PR --- FIZZY WJ --- FIZZY BB --- FIZZY BS --- FIZZY OP --- ARK PR --- ARK WJ --- ARK BB --- ARK BS --- ARK classDef coordinator fill:#1e3a5f44,stroke:#2563eb,stroke-width:2.5px classDef agent fill:#13302e33,stroke:#14b8a6,stroke-width:1.5px classDef tool fill:#27201a33,stroke:#d97706,stroke-width:1px classDef airgap fill:#88133722,stroke:#fb7185,stroke-width:1px,stroke-dasharray:5 classDef hub fill:#04785733,stroke:#059669,stroke-width:2px class OP coordinator class PR,WJ,BB,BS,RA,HN,PC,RW,JZ,RV,AT agent class AMP,CDX,OC tool class RT,AC,AIR1,AIR2 airgap class FIZZY,ARK hub
Fully Connected (12)
Optimus, Prowl, Wheeljack, Bumblebee, Bluestreak, Red Alert, Hound, Perceptor, Rewind, Jazz, Rivet, Alpha Trion โ€” connected via channels and task boards.
Tool Layer
Amp, Codex CLI, OpenCode โ€” external coding tools that Wheeljack orchestrates for implementation tasks.
Air-Gapped
Ratchet and Arcee operate in isolation โ€” no cross-agent communication. Security boundary by design.
05 The Digest Pipeline

A concrete example of multi-agent coordination: turning 194+ weekly WhatsApp messages into a published blog digest.

1
Prowl

Scan WhatsApp Group

Prowl ingests 194+ messages/week from the AI tools WhatsApp group, filtering signal from noise.

2
Prowl

Extract Highlights

Preserves Indonesian slang and cultural context. Captures tool recommendations, hot takes, and community sentiment.

3
Prowl

Structure into Categories

Organizes highlights into thematic sections: new tools, community debates, tutorials shared, notable quotes.

4
Prowl

Send to Wheeljack

Structured JSON payload via sessions_send. Data format is pre-agreed โ€” no ambiguity in the handoff.

5
Wheeljack

Format, Branch, PR

Wheeljack formats the digest as MDX, creates a feature branch, and opens a Pull Request on the blog repository.

6
Optimus

Notify for Review

Optimus pings Zain with the PR link. Human reviews, edits if needed, and merges. The digest goes live.

06 Human-in-the-Loop Philosophy

Not everything should be automated. The system explicitly tags what agents can handle autonomously vs. what needs Zain's judgment.

bot-actionable
  • Code formatting & linting fixes
  • PR creation from structured data
  • Data extraction & summarization
  • Calendar event creation
  • Routine card updates & status changes
  • WhatsApp digest compilation
  • File organization & knowledge updates
needs-human
  • Publishing blog posts or public content
  • Financial decisions & transactions
  • Architecture & design choices
  • External communications on Zain's behalf
  • Deleting or archiving important data
  • Approving agent access changes
  • Merging PRs to production
The key insight: agents should maximize preparation and minimize decision burden. A PR is ready to merge with one click. A digest is formatted and proofread. The human makes the call, not the busywork.
07 Compound Engineering

Post-work rituals that turn completed work into searchable knowledge, reusable patterns, and automated workflows.

Compound Capture ๐Ÿงช

After PR merges or bug fixes, captures learnings into knowledge/solutions/ with YAML frontmatter. Triggered by "Compound it ๐Ÿงช".

Fizzy Dispatcher

575-line bash script that auto-routes bot-actionable Fizzy cards to coding agents. Checks fuel levels, respects daily limits (5/day, 2/run), and spawns appropriate agents.

PATTERNS.md

Per-repo conventions file read by agents at session start. 90-day decay cycle. Wheeljack owns personal repos, Red Alert owns BTA patterns.

Swarm Skills

Parallel issue processing with conflict-safe wave grouping:

  • amp-swarm โ€” Amp deep mode โ†’ PR per issue
  • claude-swarm โ€” claude --print โ†’ PR per issue
  • Both feed into multi-reviewer pipeline
Multi-Reviewer Pipeline

4 parallel Opus reviewers per PR:

  • ๐Ÿ”’ Security reviewer
  • โšก Performance reviewer
  • ๐Ÿ“ Pattern compliance reviewer
  • ๐Ÿ—๏ธ Architecture reviewer

Findings classified as P1 (block) / P2 (fix) / P3 (note)

The compound engineering loop: Plan โ†’ Work โ†’ Review โ†’ Compound. Every completed task feeds back into the knowledge base, making the next task faster and more informed.
Related Pages
Shared Conventions Operations Dashboard Memory Management
Project Transformers โ€” Architecture ยท Updated 2026-03-30 17:05 WIB Blog Post  ยท  All Visualizations  ยท  The Ark