The Quality Forge | Agentic Quality Engineering

The Forge Philosophy

⚡

From Testing to Orchestration

Quality isn't tested in—it's built in. We're moving from testing-as-activity to agents-as-orchestrators, bridging classical QE with agentic intelligence through PACTS principles.

🔨

Practitioner First

Everything here is battle-tested in production. Real implementations, actual failures, honest lessons. No ivory tower theory—just what works (and what doesn't) from someone in the trenches.

🌍

Community Driven

Building the Agentics Foundation community in Serbia and sharing globally. Learning happens together—through meetups, open source, and honest conversations about what quality means in the AI age.

From the Forge

View all 35 articles →

● NEW • V3 Journey • Delivery

When the Pipeline Became the Bottleneck

Two months of parallel agents, a delivery pipeline I had to simplify, a QE organism finding its way into daily work, and a live meetup where a couple of minutes of exploratory testing found two bugs after an agent repaired 83 drifted tests.

Published September 30, 2026 8 min read

● Published • V3 Journey • Vendor Independence

The Orchestra Keeps Playing

Two weeks of hardening a platform at Ruv's pace, six podcast episodes, a first-ever Vienna meetup, four releases about honesty, a website audit that turned on its own author, and the fortnight a musician left the orchestra without the music stopping.

Published August 2, 2026 13 min read

● Published • V3 Journey • Adversarial Review

The Court That Blocked My Release

Two weeks of building quality nets around a platform racing toward release, a trip to Munich, a benchmark where the frontier model never won a task, and the moment my own AI reviewers ruled against me — QE-Court's first case was its own release, and the verdict was BLOCK.

Published July 19, 2026 14 min read

● Published • V3 Journey • Local-First AI

The Loop Came Home

Three weeks of beta-testing Reuven Cohen's MetaHarness, benchmarks proving cheap local models can generate tests without losing quality, two meetups in two formats including our first panel, and the week Adam and Klara came to Novi Sad.

Published July 5, 2026 17 min read

// AQE Fleet v3.14 - 60 Agents • 13 DDD Domains
const fleet = {
  // Queen Coordinator: Hierarchical Orchestration
  queen: "qe-queen-coordinator",  // Fleet-wide coordination

  // 13 Bounded Contexts with Specialized Agents
  domains: {
    "test-generation":     ["test-architect", "tdd-specialist"],
    "test-execution":      ["parallel-executor", "retry-handler"],
    "coverage-analysis":   ["coverage-specialist", "gap-detector"],
    "quality-assessment":  ["quality-gate", "risk-assessor"],
    "defect-intelligence": ["defect-predictor", "root-cause-analyzer"],
    "security-compliance": ["security-scanner", "security-auditor"],
    "contract-testing":    ["contract-validator", "graphql-tester"],
    "visual-accessibility":["visual-tester", "accessibility-auditor"],
    "chaos-resilience":    ["chaos-engineer", "load-tester"],
    "learning-optimization":["learning-coordinator", "pattern-learner"],
    "code-intelligence":   ["kg-builder", "dependency-mapper"],
    "requirements":        ["requirements-validator", "bdd-generator"],
    "enterprise":          ["soap-tester", "sap-rfc-tester"]
  },

  tdd: ["tdd-red", "tdd-green", "tdd-refactor", "code-reviewer",
        "integration-reviewer", "performance-reviewer", "security-reviewer"]
};

// Key: TinyDancer routing • ReasoningBank • Dream Cycles
// 9 RL algorithms • HNSW vector search • Trust Tiers

v3.14.6 • TypeScript + Node.js 22+ • MIT License

AQE Fleet: 60 Agents, 13 Domains

Domain-Driven Design architecture with 60 specialized agents across 13 bounded contexts. Queen Coordinator for hierarchical orchestration, TinyDancer 3-tier model routing, and ReasoningBank learning with Dream cycles.

TinyDancer 3-Tier Routing

Haiku/Sonnet/Opus intelligent routing with confidence-based escalation and cost optimization

9 RL Algorithms + ReasoningBank

Q-Learning, SARSA, DQN, PPO, A2C, DDPG, Actor-Critic, Policy Gradient, Decision Transformer with HNSW vector search

75 QE Skills with Trust Tiers

46 Tier 3 verified skills + 29 additional. QCSD swarms, n8n testing, enterprise integration, coherence verification

View on GitHub Read the V3 Story

LionAGI QE Fleet: Python Orchestration

Python reimplementation of the AQE framework using LionAGI orchestration. 18 specialized agents with async-first architecture, alcall integration, and 82% test coverage.

LionAGI Native Integration

Builder pattern and Session management for persistent multi-agent coordination

99%+ Reliability

alcall integration with exponential backoff and fuzzy JSON parsing (95% error reduction)

Async-First Architecture

Real-time progress streaming and parallel execution with <1ms tracking overhead

View on GitHub Read the Story

# LionAGI QE Fleet v1.2.0 - Python Implementation
from lionagi import Branch, Session

fleet = {
  "core": [
    "test_generator",      # ReAct reasoning loops
    "test_executor",       # Async parallel execution
    "coverage_analyzer",   # AST-based analysis
    "quality_gate",        # ML validation
    "quality_analyzer",    # Multi-tool integration
    "complexity_analyzer"  # Cyclomatic < 10
  ],
  "performance": [
    "performance_tester",  # Load testing
    "security_scanner"     # SAST/DAST
  ],
  "strategic": [
    "requirements_validator",  # INVEST + BDD
    "production_intelligence", # Incident replay
    "fleet_commander"          # Orchestration
  ],
  "advanced": [
    "regression_risk",     # ML test selection
    "test_data_architect", # Distributed data
    "api_contract",        # Breaking changes
    "flaky_hunter"         # Statistical detection
  ],
  "specialized": [
    "deployment_readiness", # Risk assessment
    "visual_tester",        # UI diff
    "chaos_engineer"        # Fault injection
  ]
}

# Key Metrics: 80% cost savings • 82% coverage
# pytest, Jest, Mocha, Cypress • Python 3.10+

v1.2.0 • Python 3.10+ • LionAGI Framework • MIT License

Sentinel: Multi-Agent Testing

Open-source agentic testing framework built with Rust and Python. Specialized agents working in concert— functional testing, security injection, performance planning—all with explainability first.

PACTS-Based Architecture

Proactive, Autonomous, Collaborative, Targeted, Structured from the ground up

Explainable Actions

Every agent decision comes with reasoning traces

Human-in-the-Loop

Critical checkpoints, not blind automation

View on GitHub

// Sentinel Agent Configuration
{
  "swarm": {
    "agents": [
      {
        "type": "functional-positive",
        "role": "happy_path_validator",
        "autonomy": 0.8
      },
      {
        "type": "functional-negative",
        "role": "chaos_explorer",
        "autonomy": 0.9
      },
      {
        "type": "security-injection",
        "role": "threat_simulator",
        "autonomy": 0.7
      }
    ],
    "orchestration": "hybrid",
    "explainability": "required",
    "human_checkpoints": ["pre-deploy", "security"]
  }
}

v0.1.0-alpha • Experimental archive • MIT License

Meta-memory / learning infrastructure

Nagual QE: memory that reality can correct

A Rust-native, local-first knowledge system that captures reusable patterns, retrieves them in context, and updates their quality from real outcomes. It is the portfolio's meta-memory layer—not another test runner.

Living hypotheses

Patterns carry confidence, reward, reuse, and classified failure outcomes.

Hybrid retrieval

Full-text and neural search surface relevant prior learning.

Self-improvement

Decay, consolidation, surprise, and recommendations keep memory useful.

Local ownership

Rust, SQLite/SQLCipher, optional PostgreSQL, backup, and sync.

View on GitHub

The learning loop

01
CaptureStore the problem, solution, context, domain, and provenance.
02
RetrieveBring relevant patterns back when a similar decision appears.
03
ObserveRecord success or a classified failure after the approach is used.
04
ImproveStrengthen useful patterns, expose weak ones, and consolidate carefully.

Reproducible findings and explicit limits

Experiment ledger

Published observations from Agentic QE, Nagual, MetaHarness evaluation, adversarial release work, and live exploratory sessions. Each result links to its context; none is a universal promise.

Judge qualification · controlled test

A 90% judge with zero recall was rejected

A synthetic judge scored 90% aggregate correctness while missing every evidence-corruption case. Qualification by fault slice rejects it: a healthy average cannot buy authority over the category it never catches. A test of the rule, not a field benchmark of a model.

Read the judge-qualification case →

Suite vs exploration · observed live

83 test repairs, then two bugs in a couple of minutes

At a public meetup, a sub-agent repaired 83 tests broken by drift. A couple of minutes of manual exploratory testing then found two product bugs the suite had not established. One session on one app, recorded with its rough edges.

Read the meetup account →

Cost-Pareto · observed

Cheap-first routing beat the default

Across seven QE tasks and three routing policies, the cheap tier matched or beat the frontier tier on every task at 3–4× lower cost. The session cost $0.42; the sample is deliberately small.

Read methods and caveats →

Local semantic judge · observed

On-device pattern quality reached AUROC 0.997

On a 150-pattern Nagual set, qwen3:8b and gemma4:12b-mlx separated useful knowledge from lifecycle noise locally, with no API call. This is one curated dataset, not a general model ranking.

Read the local-memory experiment →

Adversarial release gate · observed

A green suite still hid three release blockers

QE-Court's first case reviewed its own Agentic QE release and returned BLOCK after finding three data-loss or crash defects behind a passing unit suite.

Inspect the release case →

Meta-memory · ongoing

More stored patterns did not mean better memory

Nagual crossed 300 accumulated patterns, then shifted the question from volume to lifecycle quality: retrieval, outcome evidence, decay, pruning, and whether learning survives real integration paths.

Follow the learning investigation →

Join the Community

Growing the Agentics Foundation community in Serbia and beyond. Monthly meetups, open discussions, and learning together about quality in the age of intelligent agents.

Agentics Foundation Global Meetups

Join the worldwide Agentic QE community. Monthly meetups happening across the globe—from Novi Sad to San Francisco, building the future of quality engineering together.

View Global Meetup Calendar → Novi Sad Chapter Community (Members Only) →

Global Agentics Foundation

Member of the global Agentics Foundation. Building chapters worldwide, bringing PACTS principles and agentic engineering to quality practices across continents.

Explore Agentics.org →

Speaking & Workshops

Looking for a speaker on Agentic QE, PACTS principles, or bridging classical to modern quality practices? Let's talk about bringing practical insights to your conference or team.

Get in Touch

Questions about Agentic QE? Want to discuss consulting or speaking opportunities? Let's connect.

Stay Sharp in the Forge

Weekly insights on Agentic QE, implementation stories, and honest takes on quality in the AI age. No spam, no vendor pitches—just practitioner-to-practitioner learning.

Where Quality is Forged, Not Just Checked