Overview
Goal: decouple income from hours worked. AI agents run a portfolio of projects 24/7; the operator directs, the agents build, review, and ship. Four layers govern eight projects across five machines. Expand any card below for the full detail. Status is baked as of 14 Jun 2026.
The 4 layers
Each layer is a department. Top sets rules, bottom does the work. Click to expand.
L1 · Sandbox — "the rulebook" live
The constitution + safety gates every AI agent obeys. Governs everything below; no app runs here.
- CLAUDE.md — operator constitution (≤540 lines · v2.7) + verbatim aliases AGENTS.md / GEMINI.md
- audits/ — §13 install gate (allowlist + audit log) — every dependency reviewed first
- local-marketplace/ — ~88 reusable skills (the agents' procedural memory)
- Risk tiers R0 / R1 / R2 · bankai 3-agent gate on R0 actions
L2 · Hub — "the manager" building
Routes each task to the right AI model, lets projects delegate to each other, watches health.
- l2-hub :8100 + Redis :6380 · 14+ role agents · swarm bus (parallel / debate / reactive)
- LiteLLM :4000 → OpenRouter (cloud, ZDR) + local Ollama fleet (8 models)
- Strategy E routing: privacy zone → task type → escalate cheap→smart
- C.7 notify + command bus — 12/12 built (Discord + Telegram alerts · 135 tests)
- Health monitor: 8 /health/* endpoints + Uptime Kuma
L3 · SBOS — "the research brain" live
Scans the world 24/7, scores what matters, feeds the engines. Runs as Docker on the 5080.
- sbos-api :8200 · harvester 109+ sources · 6h cron · multi-axis scorer
- Qdrant 6 collections: 4,153 articles + 89 idea registry + Obsidian mirror (2,138 notes)
- Embeddings = CPU FastEmbed (bge-small, 384-dim) — no GPU needed
- Morning brief + MANIS weekly digest → Telegram · publishes /intel/summary to engines
L4 · Engines — "the product teams" 8 · mixed
The actual products. Each runs semi-autonomously, queries L3 for context. Full detail in the 8 engines.
The 8 engines
Layer 4 products. Each card = what it does, port, status, tests.
TMAO — crypto trading blocked
- What: AI analyses market, executes via ccxt. FastAPI :8009 · Next.js :3001
- Blocked: paper win-rate 33.96% — gate needs ≥70% before live capital
- OMEGA 4-engine scoring · TimescaleDB · the revenue long-pole
SBOS — intelligence brain live
L3 itself — 24/7 harvest + score. :8200 · 329 tests · 4,153 articles.
AKASHA — knowledge ops live
Gamified skill-tree / knowledge engine. :8600 · 281 tests · Waves 1-4 done.
Synapse — lab management live
RAG lab assistant (Phra Khao) over LINE OA. :8500 · 218 tests · Phase 6 complete · public via Cloudflare Tunnel api.synapse.inlife.dev.
PlatformIO — ESP32 firmware live
4 hardware repos: SpotWelder · JumpStarter · HybridPack · Zircon UPS (supercapacitor / battery PoCs · INA260 sensing).
VISTEC — MEng research active
Battery science papers. NMC90 tropical-field manuscript v8 in revision (4 co-authors).
HESS — hybrid energy storage design
Battery + supercapacitor pack design. :8300 · 3 patents · solar power-box. Lab-stage.
PMOS — media OS early
Faceless / Thai YouTube content pipeline (Whisper + Remotion). :8650 · Ch-1 Mr. Deadline pre-launch.
The 5 machines
Where everything runs. Connected by a private Tailscale network. (Phase 3 = pinning each to its machine.)
Agent protocols (6-layer)
How agents talk to tools, each other, and the world.
| Layer | Protocol | Purpose | Status |
|---|---|---|---|
| 1 | MCP | tool + data integration | live |
| 2 | A2A | agent-to-agent delegation | partial |
| 3 | AG-UI | live execution streaming | planned |
| 4 | A2UI | agent-generated UI | planned |
| 5 | AP2 | agent payments | blocked (R0) |
| 6 | x402 | HTTP micropayments | blocked (R0) |
Work pipeline
4 hard gates: a feature must be describable in 1 sentence → plan complete → all tests green → zero review issues. No code before an approved plan; nothing hits the stable branch without passing tests + review.
AI model routing
Match the model to how costly a wrong answer is — not to task size.
| Tier | Use for | If wrong |
|---|---|---|
| Haiku (cheap) | file ops, formatting, boilerplate | harmless |
| Sonnet (default) | standard dev, code review, tests | recoverable |
| Opus (premium) | architecture, trading, safety, finance | damaging |
Plus privacy zones: 🔴 secret → local Ollama only · 🟡 internal → ZDR cloud · 🟢 public → any.
Roadmap — 10 phases
| Phase | What | Status |
|---|---|---|
| 0 · Audit | 8 engines audited | done |
| 1 · Urgent fixes | safety + secrets | partial |
| 2 · Shared infra | Ollama / Qdrant / Postgres / kanban | in progress |
| 3 · Per-engine pin | spread onto 5 machines | Jun 15+ |
| 4–5 · Automation | mode switcher · execution agent | Jul |
| 6 · Fine-tune | 3 small local AI models | Jul |
| 7 · Verify | full audit green | Aug |
| 8 · Doctrine lock | apply rulebook updates | Aug |
| 9 · Productize | templates · consulting | Q4 2026 |
| 10 · Scale | SaaS · B2B | 2027 |
Port registry
Every service has a fixed home (stable / dev ports).
| Service | Stable | Service | Stable |
|---|---|---|---|
| l2-hub | 8100 | SBOS | 8200 |
| TMAO API | 8009 | HESS | 8300 |
| Synapse | 8500 | AKASHA | 8600 |
| PMOS media | 8650 | LiteLLM | 4000 |
| Redis | 6380 | Qdrant | 6333 |
| Postgres | 5432 | Ollama | 11434 |