Mini-Projects

Each architecture is accompanied by a hands-on Jupyter notebook (or Python script) that demonstrates the pattern with a practical real-world use case.

Running the Notebooks

This project uses uv for dependency management.

# Install dependencies
uv sync --group dev

# Launch Jupyter
uv run jupyter notebook mini-projects/

Part I — Core Patterns

#	Notebook	Architecture	Description
1	Smart Trip Budget Planner	Single Agent (ReAct)	A ReAct agent plans and prices a custom travel itinerary using web search and calculation tools
2	Blog Post Refiner	Prompt Chaining	Sequential LLM steps draft, critique, and polish a blog post
3	Customer Support Router	Routing	Classifies customer queries and routes them to specialized support agents
4	Editorial Review Board	Parallelization	Multiple specialist reviewers analyze a document in parallel
5	Competitive Intelligence Report Generator	Orchestrator-Worker	Orchestrator dynamically dispatches research workers to gather competitive intel
6	Data-Driven Market Analyzer	Supervisor	Supervisor coordinates research and analysis agents for market reports
7	Self-Correcting Code Generator	Reflection	Generator-critic loop writes, tests, and iteratively fixes code
8	Audience-Adaptive Content Optimizer	Evaluator-Optimizer	Adapts content to target audiences with scored quality iteration
9	Customer Service Swarm	Swarm	Specialist agents hand off conversations based on customer needs
10	Full-Stack Project Builder	Hierarchical Teams	Research and engineering teams collaborate to build a full-stack application

Part II — Quality, Verification & Oversight

#	Notebook	Architecture	Description
11	SQL Generator with Execution Verifier	Generator-Verifier	Generates SQL queries and verifies them by actual execution against a SQLite database
12	LLM-as-a-Judge Evaluator	Agent-as-a-Judge	Uses an LLM judge to evaluate and rank multiple candidate responses against a rubric
13	Expense Approval with Human-in-the-Loop	Human-in-the-Loop	Routes expense requests through automated checks and human approval gates

Part III — Parallelism, Pipelines & Cost Optimization

#	Notebook	Architecture	Description
14	Long Document Summarizer	Map-Reduce Agents	Maps chunks of a long document to summarizers then reduces to a final summary
15	Mixture of Agents Analysis	Mixture-of-Agents	Aggregates diverse model outputs via a synthesizer for higher-quality answers
16	DAG Data Pipeline	DAG / Graph Orchestration	Executes a data processing pipeline with explicit dependency ordering
17	Research Report Plan-and-Execute	Plan-and-Execute	Creates an explicit research plan then executes each step, replanning on failure
18	Cost-Optimized Model Cascade	Cascading Agents	Routes queries through cheap models first, escalating to expensive ones only when needed

Part IV — Multi-Agent Collaboration

#	Notebook	Architecture	Description
19	Blackboard Pattern	Blackboard	Specialist agents collaborate by reading and writing to a shared blackboard
20	Market-Based Task Allocation	Market-Based / Bidding	Agents bid for tasks based on their capabilities, with the best bid winning
21	Contract Net Protocol	Contract Net Protocol	Manager broadcasts tasks, contractors bid, and the best contractor is awarded the contract
22	Debate / Adversarial Collaboration	Debate / Adversarial	Two agents debate opposite positions with a judge synthesizing the final verdict
23	Red-Team Agent	Red-Team Agent	Red-team agent attacks outputs to find vulnerabilities; blue-team agent defends
24	Role-Based Collaboration	Role-Based Collaboration	Persona-driven agents with defined roles collaborate on a shared task pipeline
25	Conversational Multi-Agent	Conversational Multi-Agent	Multiple agents engage in open dialogue to solve a problem through emergent discussion
26	Event-Driven Multi-Agent	Event-Driven Multi-Agent	Agents react to asynchronous events via an event bus with no central coordinator

Part V — Reasoning Patterns

#	Notebook	Architecture	Description
27	Inner Monologue	Inner Monologue	Agent uses hidden scratchpad reasoning before producing clean final output
28	Speculative Execution	Speculative Execution	Agent speculatively executes multiple branches in parallel, selecting the best result
29	Skeleton of Thought	Skeleton of Thought	Generates an outline skeleton first, then elaborates each section in parallel
30	ReAcTree	ReAcTree	Explores multiple reasoning strategies as a tree, selecting the best branch

Part VI — Domain Applications

#	Notebook	Architecture	Description
31	Agentic RAG	Agentic RAG	Dynamically reformulates queries, routes to multiple sources, and synthesizes retrieved knowledge
32	Agentic Coding	Agentic Coding	Writes, executes, and iteratively repairs code until tests pass
33	Self-Tooling Agent	Self-Tooling Agent	Agent creates its own tools at runtime when existing tools are insufficient
34	Dynamic Tool Generation (ToolFactory)	Dynamic Tool Generation	Systematically generates, validates, and registers new tools via a factory pipeline

Part VII — Cognitive Architectures

#	Notebook	Architecture	Description
35	Neuro-Symbolic Agent	Neuro-Symbolic Agent	Combines LLM reasoning with a symbolic rule engine for verifiable logic
36	Dual-Paradigm Framework	Dual-Paradigm Framework	Routes simple queries to fast neural paths and complex ones to symbolic reasoning

Part VIII — Advanced Planning & Search

#	Notebook	Architecture	Description
37	Language Agent Tree Search (LATS)	LATS	Uses Monte Carlo Tree Search with LLM evaluation to explore solution paths
38	Introspective MCTS	Introspective MCTS	MCTS where the agent introspects and critiques each simulation step
39	Reflective MCTS	Reflective MCTS	MCTS with cross-episode reflection that improves search heuristics over time
40	Collaborative Tree Search	Collaborative Tree Search	Multiple agents explore different branches of a solution tree in parallel
41	Beam Search for Agents	Beam Search	Maintains top-K candidate solutions at each step, pruning low-scoring beams

Part IX — Agent Protocols

#	File	Architecture	Description
42	MCP Server	Model Context Protocol	Implements an MCP server exposing tools via the standard JSON-RPC protocol
43	A2A Protocol	Agent-to-Agent Protocol	Demonstrates cross-vendor agent communication using the A2A protocol
44	Agent Communication Protocol (ACP)	ACP	REST-based lightweight agent messaging using the ACP standard
45	Agent Network Protocol (ANP)	ANP	Decentralized agent discovery and communication using DID-based ANP

Part X — Agent Memory System

#	Notebook	Architecture	Description
46	Episodic Memory	Episodic Memory	Agent stores and retrieves past interaction episodes to improve future responses
47	Semantic Memory	Semantic Memory	Agent maintains a vector-based knowledge store of facts and concepts
48	Procedural Memory	Procedural Memory	Agent learns and recalls how to perform tasks from prior successful runs
49	Agentic Memory (A-MEM)	Agentic Memory	Zettelkasten-style interconnected notes with dynamic linking and retrieval
50	Collaborative Memory	Collaborative Memory	Multiple agents share a common memory store, contributing and retrieving knowledge
51	Contextual Experience Replay	Contextual Experience Replay	Agent replays high-value past experiences to improve current task performance
52	Meta-Memory	Meta-Memory	Agent manages its own memory system, deciding what to store, forget, or consolidate

Part XI — Agent Safety & Resilience

#	Notebook	Architecture	Description
53	Guardrail Agent	Guardrail Agent	Constitutional filter that validates inputs and outputs against safety rules
54	Circuit Breaker	Circuit Breaker	Prevents cascading failures by opening the circuit after repeated service failures
55	Saga Pattern	Saga Pattern	Manages multi-step workflows with compensating actions to rollback on failure
56	API Gateway / Gatekeeper	API Gateway	Centralized validation and rate-limiting layer for all agent tool calls
57	Least-Privilege Ephemeral Identity	Least-Privilege Identity	Issues short-lived, scoped credentials per agent task to minimize blast radius
58	Dead Letter / Escalation	Dead Letter	Routes failed tasks to a dead-letter queue with automatic escalation

Part XII — Agent Infrastructure

#	Notebook	Architecture	Description
59	Agentic Mesh	Agentic Mesh	Service mesh for agent-to-agent communication with discovery and load balancing
60	Agent Registry and Discovery	Agent Registry	Central registry where agents register capabilities and discover other agents
61	Agent Control Plane	Agent Control Plane	Governance layer for agent lifecycle, policy enforcement, and observability
62	Data Flywheel	Data Flywheel	Captures agent interaction data to continuously fine-tune and improve models

Part XIII — Embodied / Physical

#	Notebook	Architecture	Description
63	VLA Models	VLA Models	Vision-Language-Action model that maps visual observations to robot actions
64	Embodied AI	Embodied AI	Agent that perceives a physical environment and takes grounded actions
65	World Models	World Models	Agent uses an internal world model to simulate outcomes before acting

Part XIV — Frameworks & Meta-Approaches

#	Notebook	Architecture	Description
66	CoALA	CoALA	Uses CoALA taxonomy to compare and analyze different agent architectures
67	AEGIS Framework	AEGIS	Layered safety composition combining guardrails, monitoring, and intervention
68	ADAS	ADAS	Meta-agent that searches for and designs new agent architectures automatically
69	Self-Evolving Agents (MASE)	Self-Evolving Agents	Agent modifies its own prompts and strategies based on performance feedback
70	Nested Learning	Nested Learning	Multi-timescale optimization with inner, middle, and outer learning loops