Skip to content

🤖 AI Agents Module

Architecture, orchestration, tool-use loops, production guardrails, and interview preparation

Overview

AI agents are systems in which an LLM decides, in a loop, which tools to call and when to stop, rather than following a fixed code path. This module covers the engineering of production-grade agent systems — from single-agent ReAct loops to multi-agent coordination, MCP integration, and production deployment.

User Goal → Agent (Think → Act → Observe) → Tools/MCP → Result

Contents

# Document Description
1 Agent Fundamentals Core architectures (ReAct, Plan-and-Execute, Orchestrator-Worker), memory, tool-use patterns
2 Agent Interview Questions Staff-level Q&A with evaluation rubric
3 Agent Implementation Guide Building agents: tool loops, tool registries, MCP integration
4 Agent Production Architecture Deployment, guardrails, cost management, trade-offs
5 LangGraph Notes Graph-based agent state machines, checkpointers, human-in-the-loop
6 Python Async & Await asyncio concurrency for agents: tasks, timeouts, cancellation
7 Agent Observability Tracing, metrics, OpenTelemetry GenAI conventions, evals in production
8 Multi-LLM Architecture Routing, cost management, fallback chains across providers
9 Agent Deployment on ECS Running agents on AWS ECS/Fargate: streaming, scaling, secrets
10 System, User & Assistant Roles Message roles, instruction hierarchy, prompt-injection boundaries
11 Redis Lease Distributed locking with TTL, fencing tokens
12 Search Autocorrect Misspelling handling and query correction
13 Elasticsearch Internals Shards, segments, inverted index, near-real-time search
14 Multi-LLM Production Ops Token monitoring, rate limiting, caching, retries, A/B testing, alerting
15 Agent Memory Systems Short-term, long-term and episodic memory; retrieval and compaction

Implementation

  • implementation/ — Working Python agents:
  • simple_react_agent.py — Basic ReAct loop with tool registry
  • orchestrated_agent.py — Orchestrator-Worker pattern with async execution
  • agent_with_mcp.py — Agent discovering/using tools via MCP protocol
  • common/ — Shared utilities: memory, guardrails, LLM client, tool registry

Tests

  • tests/ — Pytest test suite:
  • test_simple_agent.py — Agent loop, input/output guardrails, parsing
  • test_orchestrated_agent.py — Worker execution, multi-agent coordination

Quick Start

cd ai-engineering/agents/
pip install -r implementation/requirements.txt

# Run the ReAct agent (uses mock LLM by default)
python -m implementation.simple_react_agent

# Run the orchestrator agent
python -m implementation.orchestrated_agent

# Run the MCP-connected agent (requires mcp>=2; it launches ../mcp/servers/calculator_server.py over stdio)
python -m implementation.agent_with_mcp

# Run tests
python -m pytest tests/ -v
  • RAG Module — Knowledge retrieval for agents
  • MCP Module — Protocol for agent-tool communication