WATCHTOWER 🗼

Autonomous Self-Healing Infrastructure Agent

Built on Hermes Agent | Nous Research Hackathon 2026

---

The Problem

3am. Server down. Engineer paged. They investigate. They fix. They sleep.

Next week — same failure. Same investigation. Same fix.

The system never learned. The engineer never slept.

---

The Solution

WATCHTOWER is an autonomous infrastructure agent built on Hermes Agent that:

• Monitors infrastructure health endpoints continuously
• Detects anomalies using EWMA statistical control limits
• Diagnoses failures using Bayesian fault probability
• Writes its own recovery skills for novel failures
• Recalls those skills instantly on future encounters
• Gets measurably faster with every incident

---

Empirical Evidence

Real data. Not simulated.

Incident	Failure Type	Time-to-Fix	Method
#1	memory_leak	36.2s	new_skill written
#2	memory_leak	0.8s	recalled_skill

44× improvement. Zero human code additions between sessions.

---

Hermes Integration

Hermes Agent autonomously:

• Read and understood the entire codebase
• Patched configuration files
• Launched the monitoring system
• Executed 15+ autonomous tool calls
• Wrote recovery skills to persistent Honcho memory

Tools used: execute_code · read_file · patch · honcho_context · honcho_conclude

---

How It Works

HERMES AGENT (brain)
      ↓
WATCHTOWER (memory + learning)
      ↓
YOUR INFRASTRUCTURE (health endpoint)
      ↑
skills written back to Honcho memory

The Loop:

Observe → EWMA Detection → Bayesian Diagnosis →
Skill Exists? → Recall (0.8s) | Write New (36s) →
Update Honcho Memory → Repeat

---

The Math

EWMA anomaly detection:

EWMA_t = λ·X_t + (1-λ)·EWMA_{t-1}
UCL = μ₀ + L·σ·√(λ/(2-λ))

Bayesian fault attribution:

P(failure | symptoms) ∝ P(symptoms | failure) × P(failure)

Priors update after each confirmed incident. The agent gets smarter over time.

---

Architecture

watchtower/
├── app/                    # Flask patient server
│   ├── app.py             # Health endpoint + failure injection
│   ├── Dockerfile
│   └── requirements.txt
├── monitor/               # Intelligence layer
│   ├── health_monitor.py  # EWMA anomaly detection
│   ├── bayesian_engine.py # Fault attribution
│   └── skill_manager.py   # Autonomous skill creation
├── skills/                # Self-authored recovery scripts
├── memory/                # Persistent learning data
│   └── learning_curve.log
├── watchtower_hermes.py   # Main Hermes-integrated loop
├── dashboard.py           # Terminal dashboard
├── inject_novel_failure.py
└── docker-compose.yml

---

Run It

# Clone the repo
git clone https://github.com/YOUR_USERNAME/watchtower-hermes
cd watchtower-hermes

# Start patient infrastructure
cd app && python3 app.py

# Open Hermes Agent
hermes

# Give Hermes this mission:
# "You are WATCHTOWER. Monitor localhost:5001/health.
#  Detect anomalies. Write skills. Show learning curve."

---

Hardware

Built and demonstrated on:

MacBook Pro 2020
Intel Core i5
8GB RAM — No GPU — No cloud compute

If Hermes runs on your machine — WATCHTOWER runs.

---

Results

44×   faster after learning
3     skills written autonomously
0     human code changes between sessions
15+   Hermes autonomous tool calls per session

---

The Claim

Most agents execute tasks.

WATCHTOWER accumulates operational expertise.

The longer it runs, the more capable it becomes — not because we updated the code, but because it updated itself.

---

Built for the Nous Research Hermes Agent Hackathon 2026 @NousResearch · Hermes Agent v0.2.0