Agentic Patterns

7 architecture patterns for building production multi-agent AI systems — distilled from 18 months of solo building across 8 shipped projects.

Not framework docs. Not tutorials. Patterns: the recurring problems, the trade-offs, and the decisions that held up under real load.

Full portfolio case study with live animations: my-portfolio-github-io-beta-five.vercel.app/projects/agentic-patterns.html

---

Why this exists

Most agentic AI content in 2026 is either:

• "Here is how to use LangChain" — framework tutorials that age in 3 months
• Research papers — correct but impractical
• Vibe-coded demos — look great, break in production

This repo documents what actually worked when building real systems: a multi-agent SDLC engine, a 6-agent marketing OS, a programmatic video pipeline, and a cross-platform wellness platform. All written by one person, all shipped.

The problems are not unique. If you are building anything with multiple LLM agents, you will hit every problem documented here. These are the patterns that solved them.

---

The 7 Patterns

#	Doc	The Problem It Solves
01	DAG vs Linear Chains	Why sequential agent chains break on complex tasks and how a Directed Acyclic Graph fixes it
02	Multi-Provider LLM Routing	How to build a system that keeps working when OpenAI has an outage, your free quota runs out, or a model is deprecated
03	Reality-First Memory	Why agent memory systems drift from reality and how to keep them honest
04	GraphDB for Agent Context	Sub-token-cost codebase navigation for agents that need structural code understanding
05	RAG That Doesn't Suck	SHA-256-keyed response caching, semantic filtering, hybrid search—what to embed vs what to keep ephemeral
06	Self-Mode vs External-Mode	How to safely build an agent system that can operate on itself without foot-gunning
07	Anti-Drift	Preventing agents from hallucinating their own architecture over long sessions

Start with 01 if you are designing a new pipeline. Start with 02 if you are fighting API reliability. Start with 07 if your agent system is behaving unpredictably after working fine last week.

Visual Architecture

Five patterns — how they connect:

graph LR
    GOAL([Goal]) --> DAG

    subgraph ORC ["01  Orchestration"]
        DAG[DAG Scheduler\nKahn's topological sort\nparallel where safe]
    end

    subgraph REL ["02  Reliability"]
        LLM[LLM Router\ncircuit breaker per provider\nBedrock → OpenAI → Ollama]
    end

    subgraph MEM ["03 · 04 · 05  Memory + Data"]
        direction TB
        REAL[Reality-First Memory\nSQLite ground truth]
        GDB[GraphDB Context\ncode as graph, sub-token nav]
        RAG[RAG + Idempotency Cache\nSHA-256 keyed responses]
    end

    DAG --> LLM
    LLM --> REAL
    LLM --> GDB
    LLM --> RAG
    REAL --> OUT
    GDB  --> OUT
    RAG  --> OUT
    OUT([Reviewable artifact])

    style GOAL fill:#0f172a,stroke:#6366f1,color:#818cf8
    style DAG  fill:#1e293b,stroke:#6366f1,color:#f8fafc
    style LLM  fill:#1e293b,stroke:#818cf8,color:#f8fafc
    style REAL fill:#1e293b,stroke:#a855f7,color:#f8fafc
    style GDB  fill:#1e293b,stroke:#a855f7,color:#f8fafc
    style RAG  fill:#1e293b,stroke:#10b981,color:#f8fafc
    style OUT  fill:#0f172a,stroke:#10b981,color:#10b981

Loading

Architecture reference (all 5 patterns annotated):

Constellation map (animated — open in browser):

The visual/ folder contains visual-patterns.html — a standalone animated constellation map showing how the 5 patterns connect and reinforce each other. No dependencies, no build step. Open directly in any browser.

open visual/visual-patterns.html
# or: python -m http.server 8080 → localhost:8080/visual/visual-patterns.html

---

Quick-Start Template

Want a copy-paste scaffold before reading the docs? TEMPLATE.md is a complete multi-agent system template covering agent contracts, memory routing, reality files, security guardrails, observability, and recovery. Fill in the <PLACEHOLDERS> for your domain and you have a production-grade starting structure in under an hour.

---

Who this is for

• Architects & Senior Engineers — building multi-agent systems, designing resilient infrastructure
• Founding engineers at AI startups figuring out production architecture
• Solo builders shipping agentic products without a team
• Learners (yes, even beginners!) — follow the Beginner's Learning Path below

Beginner's Learning Path

New to multi-agent systems? Follow this sequence:

1. Read Pattern 02 — Multi-Provider LLM Routing

   • Simplest concept: why single-provider systems fail, how to build fallback chains
   • No code required, all examples are pseudocode

2. Run agentic-systems 01 — Research Agent

   • Live code that implements Pattern 02
   • Watch how Blackboard and routing work together

3. Read Pattern 01 — DAG vs Linear

   • Understand why agents need to coordinate
   • Learn Kahn's algorithm (5-minute explainer included)

4. Run agentic-systems 05 — Bug Triage

   • See DAGRunner in action (2 agents working together)

Then tackle the remaining patterns. You'll read them as a senior engineer would.

---

Not for: LLM beginners who haven't built a single chatbot yet (learn the basics first, then come back).

---

What these patterns are NOT

• Not LangChain-specific or CrewAI-specific — patterns are framework-agnostic
• Not benchmarks or performance comparisons
• Not sponsored by any LLM provider

---

Contributing

Found a better approach to one of these problems? Opened an issue? Fixed a diagram? PRs welcome. See CONTRIBUTING.md.

---

License

CC BY 4.0 — use freely, with attribution.

---

Built from experience with: Agentic-SDLC, Agency OS, ACE App Builder, KAAL Engine, MY-VIDEO, WellnessInYou + BODH.