Strands Multi-Agent Orchestrator

Multi-agent personal assistant using Strands agent-as-tool pattern with orchestration across calendar, coding, and search.

Personal Assistant Agent is a multi-agent system built using the Strands agent-as-tool pattern, where a central orchestrator intelligently routes user requests to specialized agents for calendar management, coding, and real-time web research.

👉 GitHub Repository

Personal Assistant Agent – Multi-Agent Orchestrator

Overview

This project demonstrates how to build a modular, extensible, and production-style multi-agent system capable of handling complex, multi-domain user requests within a single interaction.

The system supports:

  • Intelligent task routing across multiple agents
  • Seamless handling of multi-step, multi-domain workflows
  • Integration of local tools, LLM reasoning, and external APIs

Architecture

The system follows an agent-as-tool orchestration pattern:

Orchestrator → (Calendar Agent Code Agent Search Agent)
  • Orchestrator: Receives user prompts and decides which agent(s) to invoke
  • Calendar Agent: Handles scheduling and agenda management
  • Code Agent: Executes code, edits files, and runs shell commands
  • Search Agent: Performs real-time web research using Perplexity MCP

This design enables composability, scalability, and clear separation of concerns.

Key Features

Multi-Agent Orchestration

  • Dynamic routing of user queries
  • Supports both single-agent and multi-agent workflows
  • Handles chained tasks across domains

Calendar Assistant

  • Create, update, and list appointments
  • Retrieve daily agenda
  • SQLite-backed persistent storage

Code Assistant

  • Python REPL execution
  • Multi-file editing support
  • Shell command execution
  • Developer journaling capabilities

Search Assistant

  • Real-time research using Perplexity MCP
  • Source-aware responses
  • Docker-based tool execution

Example Capabilities

Calendar Tasks

  • Create an appointment with structured metadata
  • Retrieve agenda for a specific date
  • Update existing appointments

Coding Tasks

  • Generate Python scripts
  • Execute code interactively
  • Assist with development workflows

Research Tasks

  • Retrieve up-to-date information
  • Provide sources for verification

Multi-Domain Task

Example:

Schedule a meeting, generate a script, and perform research in a single request.

This demonstrates true agent collaboration and orchestration.

Tech Stack

  • LLM / Orchestration: AWS Bedrock (Claude), Strands Agents
  • Search: Perplexity MCP (Docker-based)
  • Backend: Python 3.13
  • Database: SQLite
  • Environment: uv
  • Containerization: Docker

Project Structure

personal_assistant.py # Main orchestrator agents/ calendar_assistant.py # Calendar agent code_assistant.py # Coding agent search_assistant.py # Search agent (MCP + Docker) calendar_tools/ create_appointment.py list_appointments.py update_appointment.py get_agenda.py config/ env_setup.py constants.py appointments.db

How It Works

  1. User submits a request
  2. Orchestrator analyzes intent
  3. Relevant agent(s) are selected
  4. Agents execute tasks independently
  5. Results are aggregated and returned

This creates a flexible and extensible agent ecosystem.

Setup

uv venv
uv sync

Run orchestrator:

uv run python -u personal_assistant.py

Implementation Highlights

  • Agent-as-tool architecture for modular design
  • Multi-agent orchestration with dynamic routing
  • Integration of local tools, LLMs, and external APIs
  • Docker-based MCP integration for search
  • Persistent storage for real-world usability

What I Learned

  • Designing scalable multi-agent systems
  • Orchestrating heterogeneous tools under a unified interface
  • Integrating LLM reasoning with external execution environments
  • Handling multi-step workflows across domains
  • Building production-style agent architectures