9 Phases. 6 Gates.
Zero Shortcuts.
How BACON-AI orchestrates complex projects from prompt to production
Pipeline DAG
Nine sequential phases gated by six quality checkpoints. Failures loop back — no phase proceeds until its gate passes.
Problem Framing"]:::phase --> G1{"Gate 1
Context Valid?"}:::gate G1 -->|Pass| P2["Phase 2
Research & Discovery"]:::phase P2 --> G2{"Gate 2
Data Sufficient?"}:::gate G2 -->|Pass| P3["Phase 3
Analysis & Synthesis"]:::phase P3 --> P4["Phase 4
Solution Design"]:::phase P4 --> G3{"Gate 3
Design Reviewed?"}:::gate G3 -->|Pass| P5["Phase 5
Implementation"]:::phase P5 --> G4{"Gate 4
Build Passes?"}:::gate G4 -->|Pass| P6["Phase 6
Quality Assurance"]:::phase P6 --> G5{"Gate 5
Tests Pass?"}:::gate G5 -->|Pass| P7["Phase 7
Documentation"]:::phase P7 --> P8["Phase 8
Delivery"]:::phase P8 --> G6{"Gate 6
UAT Ready?"}:::gate G6 -->|Pass| P9["Phase 9
Retrospective"]:::phase G1 -->|Fail| P1 G4 -->|Fail| P5 G5 -->|Fail| P6 classDef phase fill:#13131a,stroke:#6366f1,stroke-width:2px,color:#e2e8f0,rx:12,ry:12 classDef gate fill:#1e1e2e,stroke:#06b6d4,stroke-width:2px,color:#06b6d4
Phase-by-Phase Breakdown
What actually happened during the Algorithmix website modernisation project, phase by phase.
Problem Framing
Gate 1 →Parse the original prompt. Define scope: full website modernisation for Algorithmix GmbH. Identify stakeholders, success criteria, and project boundaries before any research begins.
Research & Discovery
Gate 2 →5-10 Chrome browser agents scraped 121 URLs from algorithmix.com. 4-6 research agents benchmarked 7 competitors and 9 design references. Massive parallel fan-out to gather every piece of content, metadata, and competitive intelligence.
Analysis & Synthesis
Systems architect, persona specialist, Six Thinking Hats (perspective), and TRIZ innovation agents worked in concert. Defined 5 buyer personas with journey maps. Built the content decision matrix that determined what stays, what goes, and what gets rewritten.
Solution Design
Gate 3 →Marketing, content writer, and information architect agents designed the new site structure. 23-page sitemap. Dark theme design system with German microcopy throughout. Every page template reviewed against persona needs.
Implementation
Gate 4 →
Frontend dev agents built 23 Astro pages with Tailwind CSS. Single BaseLayout.astro component. One prompt produced the full build — from layout to every individual page.
Quality Assurance
Gate 5 →V-model test pipeline: TUT (Technical Unit Test) → FUT (Functional Unit Test) → SIT (System Integration Test) → RGT (Regression Test). Page-by-page validation. Mobile and desktop checks. Build verification. Failures loop back to Phase 5.
Documentation
Report writer and diagram agent produced a 96KB German strategic report. 16 Mermaid diagrams covering architecture, workflows, and competitive analysis. 13 PDF iterations — because rendering German umlauts correctly in PDF turned out to be surprisingly hard.
Delivery
Gate 6 →PoC deployed to a protected preview URL. USB handover package prepared for physical delivery. Tomi handover documentation including CLAUDE.md, IMPLEMENTATION.md, and DEPLOYMENT.md so the client's team can continue independently.
Retrospective
CompleteSSC (Stop-Start-Continue) protocol executed. Lessons learned documented and fed back into the framework. Anti-hallucination audit corrected 14 factual issues across report versions v5 through v13. This is where BACON-AI gets smarter.
Quality Gates Are Blocking
Gates are not suggestions. A phase cannot proceed until its gate passes. Failures are loops, not exceptions.
Blocking by Design
If a gate fails, the pipeline stops. The phase loops back and retries. No shortcuts, no overrides, no "we'll fix it later."
NPSL Governance
Gates enforce SA-001 through SA-008 governance rules. These are codified quality standards, not guidelines — they have teeth.
Real-World Example
SA-001 (Optimism Detector) caught 14 hallucinated claims in the v5 report. The gate blocked publication until every claim was verified.
Governance Rules Enforced at Gates
If You Know These Frameworks...
BACON-AI borrows concepts from orchestration systems you already understand. Here is how it maps.
Apache Airflow
"If you know Airflow, BACON-AI is like..."
A sequential phase pipeline with parallel fan-out within phases. Phase 2 fans out 10+ browser agents simultaneously, just like an Airflow DAG fans out parallel tasks. Gates act as sensors — the next phase does not trigger until the sensor confirms success.
LangGraph
"If you know LangGraph..."
A state machine with quality gates as conditional edges. Each phase is a node. Gates are the edges with conditions — the state only transitions when the gate evaluates to Pass. Fail edges loop back to the originating node.
Temporal
"If you know Temporal..."
Durable workflows with retry logic and state persistence — but the workers are LLMs, not deterministic functions. Each phase is an activity. Gates are signals. Context is preserved across the entire workflow, even when individual agent context windows expire.
AutoGen / CrewAI
"If you know AutoGen or CrewAI..."
Multi-agent orchestration with role-based specialisation — but with enforced governance. In AutoGen/CrewAI, agents advise. In BACON-AI, gates block. The pipeline physically cannot proceed until quality criteria are met. Governance rules are not guidelines; they are hard stops.
The Result
Nine phases, six gates, 15+ specialised agents, and one human in the loop. Here is what the pipeline produced for Algorithmix.