Page 1 of 5

Operation Fury

Multi-Chat Parallel Agent Build for Online Optimisers - Condensed Brief - 2026-04-10

$13.5kCurrent MRR
$100k6-Month Target
5Parallel Chats
28Agents Planned
9Active Clients

What Is Operation Fury?

A parallel multi-chat agent buildout that spins up 5 simultaneous Claude Opus sessions, each building a different slice of OO's autonomous operating system. The goal: build the agency that runs itself, then scale outreach hard.

One-line strategy: Build the agency that runs itself first, then scale outreach hard. The system IS the moat.

The Team

RolePersonFocus
Founder / OperatorDonalStrategy, sales, blocker clearing
Business Partner (50/50)TomasSales calls, pipeline, client relationships
Web DevNikkaClient sites, QA, builds
SupportAngelieComms, task management, admin
AI OrchestratorAlex (Claude)Everything else
Page 2 of 5

The 5-Chat Split

Each chat is self-contained, writes its own handover doc at session start, and has no file-level overlap with the others.

ChatMissionStatusKey Output
Fury-A Foundation + Capture DONE north-star-goals.md, quotes.json (98), ai-ranking-agent spec, blocker index (F01-F26, G01-G07), contacts, this brief
Fury-B Agent Implementations RESUMING 28 agents in graduated levels (5 Level-1 first, then Level 2+3). CEO, CMO, CSO, Sales Manager, Head SEO/GBP + AI Ranking Agent
Fury-C R&D Batcave SESSION A DONE 18 personas, 10 specialists, 125 experiments seed, SAFETY.md, Airtable sync, MANTRAS.md
Fury-D Knowledge + Book Ingestion RESUMING Book RAG pipeline, WAP.org scrape, /yt-learn batch, Intel Tracker population, contact dump folders
Fury-E VPS + Heartbeat BLOCKED on F05 Hetzner CX21, heartbeat.py, Telegram bot, cron infra

Graduated Deployment (from Tony's Agent Launch)

Don't build all 28 agents at once. Graduate in levels:

LevelAgentsWhen
Level 1 (Day 1)CEO, CMO, CSO, Sales Manager, Head SEO/GBP + AI RankingFirst build
Level 2 (Day 2)CTO, CFO, 7 sales pipeline agentsAfter Level 1 validated
Level 3 (Day 3+)10 dept heads + 12 link-building agentsAfter Level 2 validated
Page 3 of 5

Blocker Registry - Critical Path

Full registry lives at system/blocker-index.md. These are the ones that matter most right now.

Top 5 Blockers by ROI

IDTitleEffortROIStatus
F01 Add ANTHROPIC_API_KEY to .env 2 min 10/10 BLOCKING ALL
F05 Hetzner CX21 provision 45 min 9/10 Donal connecting
F03 OpenRouter API key (~$5) 10 min 8/10 PENDING
F07 Locate 5 priority book PDFs 15 min 8/10 PENDING
F10 /yt-learn batch on 8 videos 20 min 8/10 BLOCKED on F01
Critical path: F01 (API key, 2 min) unblocks ~80% of everything. Do this first.

Blocker Summary

SeriesCountPurpose
F-series (F01-F12)12Fury critical path - API keys, VPS, courses, data
F-series predicted (F13-F26)14Predicted issues during parallel run
G-series (G01-G07)7Parking-lot sub-tasks, lower priority
B-series (pre-Fury)10Original TOP 10 - Bison, Instantly, DataForSEO
R-series (R01-R07)7R&D Batcave specific
D-series (D01-D12)12Dashboard V1 env vars + DNS
Page 4 of 5

Knowledge Base Plan

Three layers, two currently empty. Fury-D fills them.

LayerPathStatusPurpose
Per-persona knowledge agents/personas/*/knowledge/ EMPTY RAG for individual personas (Hormozi, Siam Kidd, etc.)
Shared business library knowledge/business-library/ EMPTY Cross-persona RAG, queryable by all agents
Per-client knowledge knowledge/clients/<slug>/ POPULATED (9 clients) Agency delivery context

Fill Order (Fury-D)

  1. 5 priority books via scrape_book.py (Hormozi x2, Cialdini, Ogilvy, Feynman) - both persona-specific and shared library
  2. WAP.org scrape - full Siam Kidd course into siam_kidd/knowledge/
  3. /yt-learn batch - 8 queued videos into Obsidian + Intel Tracker
  4. Master Book List cross-ref - buy list for missing books

Triple-Layer Storage (decision locked)

Obsidian - human-readable markdown (browsable, editable)
ChromaDB - vector embeddings (persona RAG queries)
SQLite - metadata (search, filter, track ingestion state)
All three populated by scrape_book.py in one pass.

Intel Pipeline (continuous, not on-demand)

Once VPS is live (Fury-E), the Intel agent runs on cron every 15 minutes:

Page 5 of 5

Architecture Principles

Five principles stolen from Tony's Agent Launch doc (validated against our existing architecture).

1. Graduated Deployment
Build 5 Level-1 agents first. Use for a week. Only then add Level 2. Prevents premature abstraction, saves Opus tokens, finds real bottlenecks before you scale past them.
2. Self-Awareness Loop
Every agent writes performance_metric.json at end of each run + has a /self_review method that reads the metric and adapts. Our agents read close rate, CTR, NPS, signal-to-noise.
3. Continuous Data Moat
Intel runs on cron (every 15 min once VPS is live), NOT on-demand. 288 snapshots/day. 6-month archive = irreplicable moat. This is not optional - it is the strategic advantage.
4. Handoff Protocol
"Agents think, humans sign." Use in every client pitch. The political story that survives regulatory conversations. All escalation triggers in north-star-goals.md.
5. Marketing/Growth Split (Day 2)
Tony has separate Marketing Team + Growth Team. We fold them into CMO + Sales Manager for Day 1. Split when volume justifies it. Don't prematurely add organizational complexity.

Decision Filters (all agents apply these)

  1. Revenue Brain - does this move revenue up, protect it, or save time worth money?
  2. Cashflow priority - retainer delivery > sales > systems > internal tooling
  3. 2-hour rule - anything >2hr before evidence/revenue gets challenged
  4. Eat the frog - hardest task goes first
  5. 80/20 lens - which 20% of effort gets 80% of result?

Model Routing

MarkerModelWhen
🟣OpusJudgment-heavy, ambiguous, architectural, high-stakes
🟡SonnetModerate judgment, well-bounded but not trivial
🟢HaikuRote transforms, file generation, well-specified writes
Operation Fury - Online Optimisers - Generated 2026-04-10
Plan file: .claude/plans/spicy-tinkering-ripple.md | Blocker index: system/blocker-index.md