🧵

I Forgot 18 Things in 72 Hours

I built an AI team of 11 agents. Then I audited my own system and found security holes, untested agents, and a trail of dropped conversations. Here's every single one.

Dropped Threads

Security Holes

Untested Agents

Hours Reviewed

The Pattern

How Things Get Dropped (It's Not What You Think)

When you talk to an AI assistant, it works on one thing at a time. If you send a new message while it's in the middle of something, it stops what it's doing and switches to your new request. The old task doesn't crash — it just gets quietly forgotten.

It's like texting your employee while they're building a shelf. They put down the drill, read your text, and start on the new thing. The shelf stays half-built. Nobody remembers the shelf until someone trips over it.

The Interruption Pattern (Happened 18 Times)

🤖 AI is working on Task A...

⬇

📱 You send a new message about Task B

⬇

🤖 AI pivots to Task B immediately

⬇

❌ Task A is forgotten. No error. No warning. Just... gone.

AI starts working on security settings

Offered to disable 48 unused features. Waiting for your answer...

You message: "Oliver is broken, can you check?"

Valid request! But the AI drops the security task and switches to Oliver.

AI fixes Oliver, discovers more issues

Now working on Oliver. The security task is gone from memory.

You message: "What's this dashboard thing?"

New topic again. Oliver fix is done but security is still forgotten.

Hours later: security settings still untouched

Nobody remembers. No error message. No to-do list. Just a gap.

Security

3 Security Holes I Found in My Own System

These are real vulnerabilities that existed for hours because the conversation moved on before they were fixed.

🔑 API Key Stored in Plain Text

My AI service password (called an API key — basically a password that lets you use the AI service) was sitting in a config file that anyone on the computer could read. If someone accessed my machine, they could use my paid AI account and drain my balance.

FIXED — File locked to owner-only

🚫 No Login Attempt Limits

My agent system had no limit on login attempts. Someone could try thousands of passwords per second without getting blocked. Like a front door with no deadbolt — the lock exists but nothing stops you from trying every key.

FIXED — 10 attempts per minute, 5-min lockout

🔧 48 Unused Features Still Active

My agent system came with 51 built-in features (called "skills"). I only use 3. The other 48 are doors that don't need to be open. More open doors = more ways for something to go wrong. The AI offered to shut them down, I got distracted, and they stayed open for 12+ hours.

NOTED — Reviewing which to disable

The Lesson

Security fixes should NEVER be "I'll get to it later." They should be "stop everything, fix this now." I got lucky — nothing happened. But 12 hours with an exposed password and no login limits? That's how accounts get compromised.

System Gaps

9 Agents Created But Never Tested

I registered 11 AI agents. I tested 2 of them (MAYA and OLIVER). The other 9 exist on paper but have never received a single message. I have no idea if they work, if their instructions loaded correctly, or if they'd crash on first use.

It's like hiring 11 employees, training 2 of them, and assuming the other 9 figured it out on their own.

❓

SCOUT, ECHO, DOLLAR, BOOST, BUILDER

Never sent a single message. Status: unknown.

❓

CLIPPER, RADAR, BRIDGE, PULSE

Never sent a single message. Status: unknown.

✅

MAYA

Tested. Working. Plain language confirmed.

✅

OLIVER

Tested. Working. Delivered a real audit.

Full Audit

All 18 Dropped Threads

Here's everything that was mentioned, promised, or started — and then forgotten when the next message came in.

1. API key in plain text config file

Exposed for 12+ hours. Anyone with file access could steal the key.

FIXED

2. No rate limiting on agent gateway

Brute-force password attacks were unmitigated.

FIXED

3. 48 unused skills still enabled

AI offered to disable them. Got interrupted. Never came back.

OPEN

4. OLIVER's security audit findings never acted on

Found 4 real issues. Report was delivered. Nobody fixed anything.

OPEN

5. Agent-to-agent "bulletin board" never built

AI offered a shared notes system. AZ asked about Discord instead.

REPLACED BY DISCORD

6. "How AI handles interruptions" — notes not saved

AZ asked to save this for content. Was explained but never written to files.

SAVED IN THIS DOC

7. Website/manual project not started

Researched, stack chosen (Astro + Starlight), but no code written.

SEASON 2

8. MAYA never got the website brainstorm idea

AZ said "send it to Maya." Never sent.

OPEN

9. Model routing never implemented

BUILDER should use Claude Sonnet, ECHO should use GPT-4o-mini. All 11 agents are on the same model.

OPEN

10. Nine agents never tested

Created, registered, never sent a single message. Could be broken.

OPEN

11. Oliver keeps disappearing from dropdown (3 times)

OpenClaw UI limitation. Agents without recent sessions vanish from the menu.

DISCORD WILL FIX

12. No fallback if AI service goes down

Originally planned dual-provider. Only one was set up. If it dies, everything stops.

OPEN

13. Dashboard installed, used once, abandoned

5 containers running for a dashboard nobody was using. Ate memory for hours.

SHUT DOWN

14. Offline AI protocol never tested end-to-end

Encrypted backup runs daily. But "go offline and use backup AI" was never rehearsed.

LATER

15. API key copy gotcha — content note

AZ asked to save a tip about API keys disappearing after first view. Saved.

DONE

16. Cheat sheet of start/stop commands

Offered as standalone file. Ended up inside a bigger HTML instead.

INCLUDED IN OTHER DOC

17. Daily backup cron job never verified

Set for 3 AM. Mac might be asleep. May never actually fire.

OPEN

18. Discord setup — agreed on, never started

Researched, confirmed as the right move. Keeps getting one more thing before we start.

NEXT UP

What I Learned

5 Rules to Prevent This

✅
Don't interrupt your AI mid-task unless it's urgent.
Let it finish. Write your thought down. Send it after. Otherwise you'll both forget what it was doing.
✅
Do regular audits. Don't trust that everything is "handled."
I went 72 hours assuming everything was fine. 3 security holes and 9 untested agents say otherwise.
✅
Security fixes go FIRST. Always. No exceptions.
It's tempting to build the fun stuff. But an exposed password doesn't care about your content calendar.
✅
AI assistants don't have a built-in to-do list.
They work on what you just said, not what you said 3 hours ago. If it matters, write it down somewhere the AI checks every session.
✅
Test everything you build. Not just the first two things.
If you set up 11 agents and only test 2, you have 2 agents. The other 9 are just hopes and dreams.

"I knew I had an army. I just didn't know half of them were sleepwalking, the front door was unlocked, and my password was taped to the wall."

— AZ, after auditing his own AI system at 1 AM

What Happened Next

What We Fixed (Immediately)

🔓

Before

API key readable by anyone. No login limits. 48 unused features open.

🔒

After

Files locked. 10-attempt limit with 5-min lockout. Audit: 0 warnings.

The remaining items? They're tracked now. Written down. Not floating in a conversation that'll get compressed and forgotten.

Next step: Set up Discord so the AI agents can actually talk to each other, and AZ can watch them work from his phone. That thread got mentioned 4 times and interrupted 4 times. Not anymore.