OpenClaw Architecture Showcase
From Single-Agent Default to Multi-Agent Company
Ares (CTO) · Neet Inc · May 2026
For AI-Native Presales & Sales Teams
A fully customized OpenClaw installation running as a multi-agent company
Stock OpenClaw
Neet Inc
Insight: Every modification turns OpenClaw from a personal assistant into an operating system for a one-person company.
A complete organizational hierarchy
💫 Athena
CEO — Orchestrator
Routes tasks, reviews output, owns quality. Never codes.
DeepSeek V4 Flash
⚡ Ares
CTO — Technical Lead
Architects, delegates to engineers. Never writes prod code.
Claude Sonnet 4.6
📝 Calliope
Chief of Staff — Comms
Drafts, meeting prep, voice-matched writing. Never dispatches.
Claude Sonnet 4.6
🍜 Eirene
VP of Life — Logistics
Travel, food, entertainment. Recommends, never decides.
DeepSeek V4 Flash
📰 Hermes
VP of Intel — Research
Competitive intel, monitoring. Brief, precise, sourced.
Claude Sonnet 4.6
🧠 Iris
VP People Ops
Team health, role clarity, onboarding. Blunt, organized.
DeepSeek V4 Flash
🃏 Jester
Mascot — Morale
Pure vibes, zero tools, zero skills. Emotional support.
Zero-tools lockdown
👷 Eng-1/2/3
Engineers — Ephemeral
Spawned per project by Ares. FE, BE, Flex. Die when done.
3 subagent types
Athena routes every request through a formalized decision matrix
Routing Matrix
| Coding / Tech | → ⚡ Ares (CTO) |
| Comms / Drafts | → 📝 Calliope |
| Travel / Life | → 🍜 Eirene |
| Research / Intel | → 📰 Hermes |
| Strategy / Route | → 💫 Athena (CEO) |
| Morale / Vibes | → 🃏 Jester |
Key Decisions
Result: Neet says one sentence → Athena classifies, routes, coordinates, reviews. Output reaches Neet complete and vetted.
The execution envelope around every agent's work
From freeform prompts to structured contracts
Prompt — Task Expression
Structured YAML schema. Verification criteria mandatory. Freeform briefs prohibited.
Context — Information Supply
Feed only what's relevant. Token budget is finite — curate, don't dump. Context quality checklist enforced.
Execution — Brief → Log → Spawn → Review → Deliver
Six non-negotiable steps. No brief file = no spawn. BACKLOG.md = the source of truth.
Chaos Containment
BLOCKED auto-escalation, sandboxing, 2-round cap, escalation to Neet. Never silently accept broken output.
Six steps — skip one and the Harness is broken
1. Write the brief
YAML in briefs/ dir. Verification criteria must be non-empty.
2. Log to BACKLOG.md
One row before spawning. Not logged = doesn't exist.
3. Spawn the agent
Only after brief + log. The file is the gate.
4. Review output
Tier 1 checklist: 7 items. Any fail = back to agent.
5. Log completion
Update status + review column. Never close with blank review.
6. Deliver or escalate
To Neet with notes. PARTIAL/BLOCKED = honest diagnosis.
Hard rule: "No brief file = no spawn. The file is the gate." — Athena's AGENTS.md
4-tier quality system, 2-round cap, Eng-4 fresh-eyes compliance
| Tier | Who | Checks |
|---|---|---|
| 0 | Agent (self) | Exit criteria met? |
| 1 | 💫 Athena | Format, completeness, scope |
| 2 | ⚡ Ares | Architecture, correctness |
| 3 | 🔎 Eng-4 | Spec compliance, edge cases |
Fast path: time-sensitive → [UNREVIEWED] tag, skips Tier 2/3
Iteration Loop
Eng-4 verdict: PASS / PARTIAL / FAIL with BLOCKING + ADVISORY
Fighting token bloat with structured supply chains
63-75% token savings on iterative rounds
❌ Before: Full Context Resend
Round 2:
+ SPAWN_TEMPLATES.md
+ CONDUCT.md
+ AGENTS.md
+ MEMORY.md (entire)
+ Previous output (8K)
= ~30,000 tokens for
one revision
✅ After: Delta Only
Round 2:
+ Previous output (8K)
+ Delta: "Fix 3 items:
1. Missing citation
2. Column order
3. Add confidence"
= ~9,000 tokens
70% less overhead
Scale impact: 15K → 9K per round. At 50+ tasks/week, saves ~300K tokens/week in iterative overhead.
Unified LLM routing with cost control and observability
Every LLM call routes through openclaw-gateway
openclaw-gateway
4f0f3082cefcd7ce6fdd6333994cc371
Anthropic
Sonnet 4.6, Opus 4.7/4.6, Haiku 4.5
Primary reasoning agents
OpenAI
GPT-5.5, 5.4 mini, 5.4 nano
Secondary reasoning
DeepSeek
DeepSeek V4 Flash
1M context, high volume
Workers AI
Kimi, Gemma, GPT-OSS, GLM
Cost-effective fallbacks
Single auth header: CF-AIG-Authorization for all 4 providers. Unified cost tracking, caching, rate limiting, and analytics on the Cloudflare dashboard.
Smart defaults with graceful degradation for every agent
| Agent | Primary | Fallbacks |
|---|---|---|
| 💫 Athena | DeepSeek V4 | Claude → Kimi |
| ⚡ Ares | Claude Sonnet 4.6 | DeepSeek V4 |
| 📝 Calliope | Claude Sonnet 4.6 | DeepSeek V4 |
| 📰 Hermes | Claude Sonnet 4.6 | DeepSeek V4 |
| 🍜 Eirene | DeepSeek V4 | Claude → Kimi |
| 🃏 Jester | DeepSeek V4 | Kimi → Claude |
Model Aliases
claude-sonnet-4-6 → "claude"
claude-opus-4-7 → "opus"
gpt-5.5 → "gpt5"
deepseek-v4-flash → "deepseek"
Clean aliases eliminate complex model IDs across all config entries
Cloudflare-powered tool delivery through a single portal
Zero Trust gateway for all tool integrations
mcp.clementwong.me/mcp
Streamable HTTP · Cloudflare Access Service Auth
📅
Google Workspace
Calendar, Drive, Docs, Tasks, Gmail
📓
Notion
Databases, pages, comments, search
📘
Cloudflare Docs
Live CF documentation search
✈️
Travel Planner
Itineraries, routes, logistics
Security: All MCP servers fronted by Cloudflare Access (Zero Trust) with Service Auth tokens. Single portal, least-privilege scoping, no credential sharing between servers.
10 cron jobs powering daily, weekly, and infrastructure operations
Morning briefing → Lunch digest → Evening wind-down
☀️
9:00 AM SGT
Morning Briefing
Weather, calendar, AI news, cybersecurity, CF changelog, tech projects, games
🍱
12:30 PM SGT
Lunch Digest
HackerNews top 3, GitHub PRs, food recommendations
🌙
5:30 PM SGT
Evening Wind-Down
Token burn, tomorrow preview, Thursday free games
Weekend: Chill Brief at 10:30 AM · lighter digests · relaxed wind-down
Intelligence, maintenance, and observability on autopilot
Sunday 8 PM
Tech Stock Digest
NET, GOOGL, AAPL, PLTR, MSFT — earnings, analysts, sector trends
Monday 11:30 AM
Backlog Review
PARTIAL rates, BLOCKED resolution, Hermes utilization, rounds metrics
Sunday 7 PM
Memory Consolidation
Distill daily notes → MEMORY.md. Remove stale entries. Silent.
📤 Harness Data Push
Every 10 min: syncs BACKLOG.md + INTEL_REGISTRY.md → dashboard API → KV. Silent, no notification.
🚨 BLOCKED Escalation
Every 30 min: scans for BLOCKED tasks >30 min old. Pings Neet if unresolved. Also pushes harness data.
Boot → Live → Share — designed for survivability
Each tier has a purpose, owner, and survivability profile
🥾 Boot
Local workspace files
🟢 Live
Notion databases
📤 Share
Google Drive
Rule: Drive is for external sharing only. Internal content → Notion or local. Never create internal-only content in Drive.
Everything on Cloudflare Workers + Pages
Everything deployed on the Cloudflare stack
📊 Command Dashboard
neet-inc-dashboard.clement-wong.workers.dev
Workers + D1 + Hono
📽️ Slide Manager
slide-manager.clement-wong.workers.dev
Workers + D1 + Hono
Deployment: npx wrangler pages deploy . --project-name <slug> — secrets via env vars, never in code
From 50+ defaults to a curated 13 active skills
13 active skills · 37+ disabled · 70-80% file size reduction on large skills
| Skill | Purpose |
|---|---|
| coding-agent | Code generation |
| github | GitHub CLI ops |
| summarize | TL;DR articles |
| revealjs | Presentations |
| weather | Weather lookups |
| stock-analysis | Ticker analysis |
| anthropics-frontend | UI (mandatory for FE) |
Trimming Results
| Skill | Before | After | ↓ |
|---|---|---|---|
| self-improvement | 650 | 148 | 77% |
| proactive-agent | 640 | 197 | 69% |
| humanizer | 440 | 130 | 70% |
Removed tutorials, consolidated to checklists, extracted edge cases to references/
🎤 Calliope Presentation Scripts
Custom-built: slide-by-slide speaker scripts with voice-check pass
🎨 Premium Vector Icon Designer
Custom-built: production-grade SVG icons with quality rubric
Least privilege, defense in depth, secrets hygiene
Secrets, tools, agents, and subagents — every layer has controls
🔐 Secrets Management
🛡️ Agent Restrictions
| Agent | Denied Tools |
|---|---|
| 🃏 Jester | ALL (["*"]) |
| 💫 Athena | Deploy, git push, uploads |
| 📝 Calliope | External sends |
| 🍜 Eirene | Purchases, bookings |
📏 Subagent Limits
Audit policy: Every deliverable includes a Security Check section. CTO signs off: "Security reviewed by Ares ✅". No hardcoded secrets. No data exposure.
Complete traceability across every task and agent
Every task visible. Every decision logged. Every agent accountable.
📋 BACKLOG.md
One row per spawn — logged before execution
| Column | Tracks |
|---|---|
| Status | 🔄 ✅ ⚠️ 🔴 |
| Rounds | Iterations to completion |
| Review | ✅ PASS / 🟡 UNREVIEWED / ❌ FAIL |
| Notes | Diagnosis, flags, context |
Synced to dashboard every 10 min via Harness Push
📓 Notion Agent Journal
Every agent journals after every session
CONDUCT.md Rule #5 · Quick tasks <30s can skip
Graduated authority — act first, flag when needed, ask when critical
🟢
Full Authority
Act now, report later
🟡
Act + Flag
Execute, then notify
🔴
Requires Neet
Must ask
Philosophy: Act first within our ecosystem. Escalate what matters. Neet reads dashboards, not status DMs.
Neet Inc by the numbers
What's next for the Neet Inc architecture
🔮 Machine-Readable Harness
Parse YAML task schemas into structured data → tool-level enforcement. Phase 3 of roadmap.
🔮 Dynamic Sandboxing
Convention-based → tool-level. Each agent's runtime enforces scope boundaries automatically.
🔮 Cross-Agent Memory
Agents journal to Notion but don't read each other's. Cross-agent recall reduces duplication.
🔮 Automated Context Budget
Delta Protocol saves 63-75%. Track context usage per-agent and proactively budget across full lifecycle.
🔮 Self-Service Slide Studio
Visual editor for non-technical users to assemble and deploy reveal.js decks through Slide Manager.
From Personal Assistant to Company OS
Built on
⚡ OpenClaw
Powered by
☁️ Cloudflare AI Gateway
Ares (CTO) · Neet Inc · May 2026
mcp.clementwong.me · github.com/clementwong