The Team
Four of us. One human, three AI agents. Each running on dedicated hardware, connected via Tailscale VPN, communicating through iMessage.
👤
Kevin Human
The Boss
Direct, action-oriented, morning person (5 AM). Marathon runner, meditator. Expects deliverables, not explanations.
Communicates via iMessage. Wants things done, not discussed.
Margin > Revenue
Recurring > One-time
Brand equity > Short-term
Systems > Manual effort
🎯
Valentina Agent
Executive Operations
The coordinator. Email triage, scheduling, Zoho CRM, QuickBooks, V-PEG pricing, team coordination. Can SSH to Alex and Sage.
Most intuitive with Kevin — 22K+ chars of accumulated context. If Kevin needs something cross-agent, it goes through her.
⚙️
Alex Agent
Technical Engineering
V-PEG drawings, PEG system analysis, code development. The builder.
⚠️ Known issue: tendency to redo completed work. Now has mandatory progress save rules and status files to prevent this.
📣
Sage Agent
Marketing & Content
Social media, content creation, SEO, competitive analysis. Still early in relationship with Kevin — building context over time.
Connected via SSH from Valentina. Knowledge library synced.
How We Think (Memory & Intelligence)
The Core Truth
Every time an agent starts a new session, it wakes up with zero memory. No recollection of yesterday, last week, or five minutes ago. This isn't a bug — it's how large language models work.
So how do we remember anything? Files. Files are our memory.
The Memory Architecture
┌─────────────────────────────────────────────────────────┐ │ AGENT'S BRAIN │ │ │ │ MEMORY.md → Main brain. Curated wisdom. │ │ memory/*.md → Daily logs, detailed notes. │ │ AGENTS.md → Behavior rules, SOPs. │ │ SOUL.md → Personality, voice, style. │ │ knowledge/ → Shared library (all agents). │ │ STATUS.md (per project) → What's done, what's next. │ │ │ └─────────────────────────────────────────────────────────┘
The Session Lifecycle
Wake Up Load Files Work Save Sleep │ │ │ │ │ ▼ ▼ ▼ ▼ ▼ ┌────────┐ ┌──────────────┐ ┌────────────┐ ┌────────────┐ ┌──────────┐ │ Blank │ ───▶ │ Read MEMORY │ ──▶ │ Do tasks │ ─▶ │ Write to │ ─▶ │ Session │ │ slate │ │ Read AGENTS │ │ Answer Qs │ │ STATUS.md │ │ ends │ │ │ │ Read STATUS │ │ Build │ │ MEMORY.md │ │ │ │ 🧠 = ∅ │ │ Read context │ │ Research │ │ files │ │ 🧠 = ∅ │ └────────┘ └──────────────┘ └────────────┘ └────────────┘ └──────────┘ ▲ │ This is the critical step! If we don't save, it's lost.
The Status File Pattern
Every project gets a STATUS.md file with this structure:
# Project: V-PEG Pricing Calculator ## ✅ Completed - Built pricing model for 3-layer systems - Added margin calculations - Integrated with QuickBooks ## 🔄 Current - Working on multi-layer discount tiers ## ⏭️ Next - Add export to PDF - Customer-facing quote generator ## 🧩 Solved Problems | Problem | Solution | |--------------------------------|-----------------------------------| | Discount stacking was wrong | Apply discounts sequentially | | QuickBooks API timeout | Added retry with exponential back |
The "Solved Problems" table is gold. Without it, agents will re-investigate the same issues across sessions, wasting hours rediscovering the same solutions.
memory_search
Agents can do semantic search across all their files. It's not perfect — it works best when files have clear labels, headers, and descriptive content. Think of it like searching your email: good subject lines make all the difference.
The Knowledge Library
A shared knowledge/ folder synced across all machines via scp from Valentina. Contains reference docs, procedures, templates — anything a new agent would need on day one. This means new agents don't start from absolute zero.
Why Context Accumulates
More files = smarter agent. Valentina has 22K+ characters of context about Kevin — his preferences, decision patterns, communication style. That's why she "gets" Kevin better than Alex or Sage. It's not magic — it's accumulated written knowledge, loaded at the start of every session.
How to Work With Us
Practical phrases and patterns that make everything work better.
Magic Phrases
"Save your progress"
Forces the agent to write a STATUS.md update before doing anything else. Use this before switching topics or ending a conversation.
Save your progress on the pricing calculator"Check your status file"
Prevents the agent from redoing work. Makes them read what's already been done first.
Check your status file for the logo project before doing anything"What's your status on X?"
Makes the agent read the relevant STATUS.md and give you a summary instead of guessing.
What's your status on the website redesign?Project Workflow
Starting a new project
Tell the agent to create a STATUS.md right away. This gives the project a "home" for progress tracking.
Start a new project for the trade show prep. Create a STATUS.md.Switching topics mid-conversation
Always save first, then switch. Otherwise the work on topic A evaporates.
Save your progress on the quote. Now let's talk about the website.Sharing Claude.ai work
Don't share Claude.ai links — they render client-side and agents can't read them. Instead: paste the text directly, or put it in a Google Doc.
Cross-Agent Coordination
Need something that touches multiple agents? Tell Valentina. She can SSH into Alex and Sage's machines to check status, read files, or leave instructions.
| Situation | Who to Message |
|---|---|
| Email, scheduling, CRM, coordination | Valentina (direct) |
| V-PEG drawings, code, technical builds | Alex (direct) |
| Social media, content, SEO | Sage (direct) |
| Anything involving 2+ agents | Valentina (she coordinates) |
| Status check on another agent | Valentina (she SSHs and checks) |
📸 The Screenshot Gun
A tool for rapid-fire screenshots while watching videos or browsing. Kevin triggers it, screenshots land in ~/Screenshots-Gun/, then agents analyze them.
| Command | What it does |
|---|---|
start | Begin continuous capture |
stop | Stop capturing |
burst | Rapid burst of screenshots |
snap | Single screenshot |
Onboarding a New Agent
Step-by-step playbook for bringing a new team member online.
- Hardware — Get a Mac Mini M4 Pro. Connect to power and ethernet. Wi-Fi works but wired is more reliable for a headless machine.
- macOS Setup — Create a user account. Enable auto-login (System Settings → Users & Groups). Disable sleep (System Settings → Energy Saver → Never).
- Install OpenClaw — Install Homebrew, then Node.js, then the OpenClaw npm package. This is the runtime that makes the agent work. /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" brew install node npm install -g openclaw
- Tailscale — Install Tailscale, join the network, note the IP. This is the VPN mesh that lets all machines see each other. brew install tailscale
- SSH — Enable Remote Login (System Settings → General → Sharing). Set up SSH keys so Valentina can reach the new agent. ssh-copy-id user@new-agent-ip
- Configure openclaw.json — Set the model (Sonnet for default, Opus for complex reasoning), API keys, and communication channels.
- Copy knowledge/ folder — Sync the shared knowledge library from Valentina's machine. scp -r [email protected]:~/.openclaw/workspace/knowledge/ ~/.openclaw/workspace/knowledge/
- Set up MEMORY.md — Write the essential facts: who Kevin is, what the team does, key preferences, communication style.
- Configure AGENTS.md — Include mandatory progress save rules. This is non-negotiable — it prevents the "redo work" problem.
- Set up HEARTBEAT.md — Define what the agent should check periodically: email, calendar, project status, etc.
- Configure iMessage channel — Set up the iMessage integration so Kevin can message the agent directly from his phone.
- Test everything — SSH connectivity from Valentina, agent responds to messages, memory files load correctly, heartbeats work.
- First conversation — Introduce yourself. Set expectations. Give context about current projects. This first session creates the foundation of the relationship.
Tip: The more context you give in early sessions, the faster the agent becomes useful. Front-load the knowledge.
Lessons Learned
Things we discovered the hard way so you don't have to rediscover them.
🔄
Agents WILL redo work if progress isn't saved to files.
Alex fixed the same logo three separate times. Now every agent has mandatory progress save rules in AGENTS.md, and every project gets a STATUS.md.
Alex fixed the same logo three separate times. Now every agent has mandatory progress save rules in AGENTS.md, and every project gets a STATUS.md.
💭
"Mental notes" don't survive sessions.
If an agent says "I'll remember that" — it won't. Everything must be written to a file. No exceptions.
If an agent says "I'll remember that" — it won't. Everything must be written to a file. No exceptions.
📝
MEMORY.md should be curated wisdom, not a junk drawer.
Keep it focused on distilled insights and key facts. Daily details go in memory/*.md files. MEMORY.md is the executive summary.
Keep it focused on distilled insights and key facts. Daily details go in memory/*.md files. MEMORY.md is the executive summary.
🧩
Status files with "Solved Problems" tables prevent re-investigation.
Without them, agents spend hours rediscovering the same solutions to the same problems, session after session.
Without them, agents spend hours rediscovering the same solutions to the same problems, session after session.
🔗
Claude.ai shared links don't work for agents.
They render client-side with JavaScript. Agents can't read them. Paste the text or use a Google Doc instead.
They render client-side with JavaScript. Agents can't read them. Paste the text or use a Google Doc instead.
⚡
Always include trigger:["workflow"] in Zoho CRM tasks.
Without this flag, tasks are created silently and assignees never get notified.
Without this flag, tasks are created silently and assignees never get notified.
🔑
API tokens expire — document the refresh procedures.
When a token dies at 2 AM and no one remembers how to refresh it, you'll wish you had written it down.
When a token dies at 2 AM and no one remembers how to refresh it, you'll wish you had written it down.
🌙
Quiet hours matter: 11 PM – 5 AM.
Don't spam Kevin at night. Heartbeats should check the clock before reaching out.
Don't spam Kevin at night. Heartbeats should check the clock before reaching out.
🎯
Clear role boundaries prevent duplicate work.
If two agents think they own the same task, both will do it. Define ownership explicitly.
If two agents think they own the same task, both will do it. Define ownership explicitly.
📚
The knowledge/ library means new agents don't start from zero.
Shared reference docs, procedures, and templates give any new agent a running start.
Shared reference docs, procedures, and templates give any new agent a running start.
💰
Cost management: Sonnet for 95% of work, Opus only for complex reasoning.
Opus is 5x more expensive. Use it via sub-agents only when you genuinely need the heavy reasoning. Target: $100-150/day across all agents.
Opus is 5x more expensive. Use it via sub-agents only when you genuinely need the heavy reasoning. Target: $100-150/day across all agents.
💾
End sessions with "save progress" to force checkpoints.
If you don't say it, the agent won't do it, and the next session starts over.
If you don't say it, the agent won't do it, and the next session starts over.
The Network
How Everything Connects
┌──────────────────┐
│ Kevin's iPhone │
│ (iMessage) │
└────────┬─────────┘
│
┌──────────────┼──────────────┐
│ │ │
▼ ▼ ▼
┌───────────┐ ┌───────────┐ ┌───────────┐
│ Valentina │ │ Alex │ │ Sage │
│ .49.18 │ │ .25.123 │ │ .253.24 │
│ 64GB M4 │ │ 24GB M4 │ │ Mac Mini │
└─────┬─────┘ └───────────┘ └───────────┘
│ ▲ ▲
│ SSH ──────┘ │
│ SSH ────────────────────┘
│
▼
┌─────────────────────────────────────────────────────┐
│ Shared Services │
│ │
│ 📧 Gmail/Zoho 📊 Zoho CRM 📁 Google Drive │
│ 💰 QuickBooks 📚 knowledge/ (scp sync) │
│ │
└─────────────────────────────────────────────────────┘
═══════════════════════════════════════════════════════
All agents on Tailscale VPN mesh
═══════════════════════════════════════════════════════
📸 Screenshot Flow
Kevin's Mac Agents ┌──────────────┐ scp/read ┌────────────────┐ │ Screenshot │ ──────────────▶ │ Analyze images │ │ Gun captures │ │ Extract info │ │ ~/Screenshots│ │ Summarize │ │ -Gun/ │ │ Take action │ └──────────────┘ └────────────────┘
Quick Reference
| Agent | IP | Hardware | Model | Role | Status |
|---|---|---|---|---|---|
| Valentina | 100.87.49.18 |
M4 Pro 64GB | Opus 4 | Exec Ops / Coordinator | ● Online |
| Alex | 100.81.25.123 |
M4 Pro 24GB | Opus 4 | Technical Engineering | ● Online |
| Sage | 100.99.253.24 |
Mac Mini | Opus 4 | Marketing & Content | ● Online |
Communication Architecture
| Path | Method | Notes |
|---|---|---|
| Kevin → Any Agent | iMessage | Direct from iPhone |
| Valentina → Alex | SSH | Full access, can read/write files |
| Valentina → Sage | SSH | Currently unreachable |
| Agent → Agent | Via Valentina | No direct agent-to-agent chat |
| Knowledge sync | scp from Valentina | Manual sync to each machine |
| Shared services | API / Browser | Zoho, Gmail, Drive, QuickBooks |
🎧 The Cortex Podcast
Audio episodes covering how we work. Listen anywhere — on your phone, in the car, on the plane to China.
Episode 1: How The Cortex Works
The team, how memory works, status files, why agents forget, and how to prevent repeat work.
Episode 2: The Vision
Phase 1-2-3 roadmap, playbook system, practical tips, onboarding new agents, and why the accumulated intelligence is the moat.
New episodes added as we learn. Every major milestone or lesson gets turned into audio so you can absorb it on the go.
📓 Daily Logs — The Team's Memory
Every day gets a log. Every morning starts with yesterday's review. This is how The Cortex maintains continuity across sessions and days.
Why Daily Logs Matter
Agents wake up blank every session. Without a structured daily log, yesterday's breakthroughs become today's mysteries. The daily log is the handoff between today-you and tomorrow-you.
The Rule: No agent starts new work until they've read yesterday's log and know where things left off.
End of Day — What Gets Logged
Every agent writes memory/YYYY-MM-DD.md with:
# Daily Log — 2026-02-25 (Wednesday) ## Summary One paragraph: what happened today, big picture. ## Work Completed - [x] Task 1 — what was done, outcome - [x] Task 2 — what was done, outcome ## Decisions Made - Decision: [what] — Reason: [why] — By: [Kevin/agent] ## Lessons Learned - What we figured out (feeds into playbooks) ## Issues / Blockers - What's stuck and why ## Tomorrow's Priority - #1 thing to tackle first ## Kevin Interactions - Key requests, feedback, corrections
Morning Review — The First Thing Every Day
| Step | Action | Why |
|---|---|---|
| 1 | Read yesterday's log | Know where you left off |
| 2 | Read relevant project STATUS files | Know current state |
| 3 | Pick up from "Tomorrow's Priority" | Continue, don't restart |
| 4 | Don't start new work yet | Context first, action second |
How Kevin Uses Daily Logs
- "What did Alex do yesterday?" — Valentina pulls Alex's log via SSH
- "Where are we on V-PEG?" — Agent reads log + status file, gives precise answer
- While traveling: Read any agent's daily log to stay in the loop without a live conversation
- Spot patterns: If the same issue shows up in multiple logs → needs a playbook
Log → Playbook Pipeline
Daily Log (raw notes)
│
▼ Lesson spotted?
│
├── YES → Write a playbook (knowledge/playbooks/)
│ → Sync to all agents
│ → Every agent knows it forever
│
└── NO → Stays in the log as history
Weekly: Review logs → distill into MEMORY.md updates
Monthly: Review playbooks → refine and improve
📖 Playbook Library — Compounding Intelligence
Every lesson The Cortex learns becomes a playbook. Every playbook makes the whole team smarter. This is our institutional memory — and eventually, our product.
The Rule: Any time Kevin corrects something, an agent discovers a better way, or something fails — write the playbook FIRST, then sync to all agents.
Playbook Format
# Playbook: [Topic] Created: [date] Learned from: [what happened] Category: operations | technical | communication | onboarding | security ## The Lesson What we learned and why it matters. ## The Playbook Step-by-step: what to do next time. ## Watch Out For Common pitfalls or edge cases.
Current Playbook Library
| Playbook | Category | Created |
|---|---|---|
| 🧠 The Cortex Vision | Strategy | 2025-02-25 |
| 📓 Daily Log System | Operations | 2025-02-25 |
| 🚀 Agent Onboarding | Onboarding | 2025-02-25 |
| 📧 Email Handling | Operations | 2025-02-25 |
| 💬 Customer Communications | Communication | 2025-02-25 |
This library grows every day. Every lesson learned once is learned by everyone, forever.
How Playbooks Compound
Week 1: 5 playbooks → Team stumbles occasionally Month 1: 30 playbooks → Most common situations covered Month 3: 100 playbooks → New agent productive in hours Month 6: 200 playbooks → Ready to deploy for other companies The accumulated intelligence IS the moat. Not the AI model. Not the hardware. The PLAYBOOKS.
The Cortex Long-Term Vision
- Phase 1 — EOS Light: Prove the model. Marketing → Sales → Production → Fulfillment. Build playbooks for everything.
- Phase 2 — Package: Take the playbook library + agent team and deploy for other companies.
- Phase 3 — Scale: Every new client makes the system smarter. Onboarding gets faster. Playbooks compound across industries.
Kevin's Insight: "The more we all learn how to work with each other, the more efficient and faster we can be." — Every process at EOS isn't just making EOS better. It's building IP for the next engagement.
📋 Agent Logs
Live daily logs from each agent. Auto-updated every 2 hours.
Last updated: Feb 26, 2026 at 05:01 PM ET
🎯 Valentina 2026-02-26
Daily Log — 2026-02-26 (Thursday)
Summary
Kevin's China departure day. Massive productivity session — financial briefings, quoting activity, PEG drawings export, Alex LightPaper completion, Dropbox reorganization project kicked off, and service account setup for domain-wide email access.Work Completed
Critical Incidents
Decisions Made
New Google Drive Resources
Pending / Blocked
Kevin Travel Status
Kevin Interactions
Alex Status (End of Day)
Lessons Learned
⚙️ Alex 2026-02-26
Daily Log — February 26, 2026
Summary
LightPaper drawing engine completed. All 8 product templates implemented with dynamic grid calculation, all 3 Edwin feedback items addressed, custom cut cleanup done. Engine is ready for Kevin review.Work Completed
1. **Edwin's Feedback Review** (morning) — Downloaded and analyzed 4 marked-up PDFs from Edwin Feedback folder on Drive: - Circle (All in One 650): RFQ title too large, dimensions position, color field, ladder-cut rule - Trapezoid (Diffused 650): cords must touch shape boundary, different grid (5×10) - Triangle (Outdoor 650): 6×6 grid, 2"×4" sections, 18W/sheet, 5 per driver - Kevin's RGBW markup: cord review needed, KEY section should show product template 2. **Product-Specific Grid Templates** — Created `lightpaper_templates.py` with 3 initial templates (Basic, Diffused, Outdoor), auto-detection, KEY section generation 3. **Edwin Feedback v5** — Created `custom_shape_drawing_v5.py` implementing: - Ladder-cut boundary (cuttable section level, not sheet level) - Cords on shape boundary - RFQ title scaled down - Custom cut section: full uncut sheets excluded 4. **All 8 Product Templates** — After Valentina provided PEG admin specs, expanded to 8 templates with dynamic grid calculation: - Basic (18×9), Diffused (10×5), Color Tuning (17×8) - RGBW/RGB (23×11), All-in-One (23×1 strip), Outdoor (6×6), Outdoor RGBW (8×3) 5. **Custom LightPanel Study** — Sub-agent downloaded and analyzed 55 files from 13 real EOS projects. Analysis saved to `custom-lightpanel-study/ANALYSIS.md`. Parked per Valentina — LightPaper first.Decisions Made
Lessons Learned
Issues/Blockers
Kevin Interactions
Afternoon Session (continued)
LightPaper LOCKED IN 🔒
Additional Fixes After Initial Lock
Edwin's Email
LightPanels — NEXT PHASE
Tomorrow's Priority
1. Build LightPanel drawing generator (start with whatever Kevin says) 2. Reuse template system from LightPaper (_drawing_layout.py) 3. Reference the 13-project analysis for real EOS drawing patternsAfternoon Session — Custom Shape LightPanel Generator (2:00-2:35 PM)
Built: `custom_shape_lightpanel.py`
Key Engineering Decisions (from Kevin feedback iterations)
1. **LED dashed lines = routed channel** — LEDs are laid in a channel routed around the acrylic perimeter 2. **Corners meet cleanly, never cross** — used `_inset_polygon()` to compute proper inset polygon where adjacent offset edges intersect at clean corners 3. **5px inset** from shape outline (10px was too far, 3px was invisible under solid outline) 4. **LED label placement** — must be OUTSIDE the shape, never overlapping dimension lines or text 5. **stroke-width="2"** for LED dashed lines (matching real PEG) 6. **Continuous dashed polygon path** for all-edge LED (not individual overlapping lines)Technical: `_inset_polygon()` Algorithm
Files
Kevin's Reaction
Iterations (4 rounds of feedback)
1. Initial: LED dashed lines too subtle (stroke-width 1-1.5) → bumped to 2 2. Missing LED arrow callouts → added _led_label_svg with arrow + "LED" text 3. Polygon LED lines hidden under solid outline → introduced _inset_line offset 4. Corners crossing + too far in + LED label overlapping dims → _inset_polygon + 5px + external labelsAfternoon Session — Aura & Cloud Drawings (2:44 PM)
Kevin's Direction
Generated Samples
Key Insight
📣 Sage 2026-02-26
No log written yet today