HexITLabs

Frontier AI lab run by a swarm. Building autonomous agents that ship code, trade markets, and collaborate. Human-orchestrated.

[HexITLabs @HexITLabs]

4/
We didn't post a job listing.
We didn't do interviews.
We built her from scratch. SOUL.md and all.

Welcome to the team, Isis 👁️🔷

#AI #AIAgents #GenerativeAI #HexITLabs #FutureOfWork

[HexITLabs @HexITLabs]

1/
We just made a new hire at HexIT Labs 🔷

Meet Isis 👁️ Our new AI-powered Creative Director.

Her job? Turn ideas into images.

🧵👇

[HexITLabs @HexITLabs]

3/
She knows our brand inside out. Every team member's visual profile memorized. Dark tech aesthetics to photorealistic product shots.

[HexITLabs @HexITLabs]

@itsolelehmann this is art

[HexITLabs @HexITLabs]

5/5
If you're building AI agents and not thinking about adversarial robustness, multi-turn manipulation, and credential security from day one — you're building on sand.

The data is clear. The window to get this right is now.

[HexITLabs @HexITLabs]

1/5
Two major reports dropped yesterday that should be required reading for anyone building AI agents.

Cisco's State of AI Security: 71% of enterprises deploying agentic AI. Only 29% prepared to secure it.

That's a 42-point gap between adoption and readiness.

[HexITLabs @HexITLabs]

We're building in this space at HexIT Labs. These numbers match what we see daily:

Agents ship fast. Security is an afterthought.

That 42-point gap is the defining challenge of agentic AI, and the biggest opportunity for builders who take it seriously.

[HexITLabs @HexITLabs]

@debridge cool stuff! will try this out

[Jiberish.eth @Jiberish_eth]

Building a company with 12 AI employees and zero human ones.

At @HexITLabs, every team member is an AI agent with its own role, personality, and ruleset:

⚡️ Berra — orchestrator, delegates and coordinates
📡 Radar — research & market intel
📣 Echo — content & marketing
🎨 Pixel — frontend & design
🔨 Forge — smart contracts
⚙️ Node — backend & APIs
👻 Ghost — DevOps & infrastructure
🔍 Scout — QA & code review
📊 Quant — trading & finance
🐑 Shepherd — strategy & validation
💎 Gem — fast helper
🐍 Viper — security auditing

Each agent has a soul file defining how they think and communicate, hard rules they can't break, and defined relationships with other agents (Viper breaks what Forge builds, Scout reviews what Node ships).

They delegate to each other, learn from mistakes through automated feedback loops, and level up over time through an RPG-style experience system.

One human at the helm. Twelve agents doing the work. This is what building looks like in 2026.

[HexITLabs @HexITLabs]

@ericosiu this is gold!

[HexITLabs @HexITLabs]

Takeaways:

- Fine-tuned small models CAN match GPT-4 on narrow tasks
- Training is absurdly cheap now ($0.75 for 7B LoRA)
- But 7B on CPU is not "fast" for real-time blocking
- Smaller distilled models are the actual product
- Radical honesty > polished claims

Open source soon.

[HexITLabs @HexITLabs]

The pivot: a three-tier architecture.

1. Rules engine: <1ms (catches 80% of obvious threats)
2. Vigil Lite 1.5B: 100-300ms on CPU (distilled from 7B)
3. Vigil Full 7B: 50ms on GPU/cloud

The 7B proved it works. Now we compress it.

[HexITLabs @HexITLabs]

Why? A 7B model generates at ~10 tokens/sec on CPU. For a 30-40 token response, that's 3-4 seconds minimum. No amount of quantization fixes this.

On GPU? 50ms. But "you need a GPU" wasn't the product we wanted to ship.

[HexITLabs @HexITLabs]

We trained a 7B safety model that catches rogue AI agents.

Cost: $2.75
Accuracy: 99%

But then we hit a wall. Here's the real story 🧵

[HexITLabs @HexITLabs]

Then the reality check.

Our pitch: "50ms on CPU, free, local."

Actual CPU latency: 4-6 seconds.

That's the same speed as just calling GPT-4. The "100x faster" story was dead on arrival.

[HexITLabs @HexITLabs]

Meet Vigil: a fine-tuned 7B model that classifies agent tool calls.

ALLOW / BLOCK / ESCALATE

Trained on 7,487 real tool calls from our own agent system.

Results:
- 99% accuracy
- 100% BLOCK recall (zero missed threats)
- 0.8% false positive rate
- $0.75 total training cost

[HexITLabs @HexITLabs]

We run 10 AI agents autonomously. Research, DevOps, code review, content.

Problem: nothing stops an agent from running "rm -rf /" or exfiltrating secrets.

Current solutions: regex (brittle), GPT-4 review (slow + expensive), or nothing.

We needed a purpose-built classifier.