DEV Community

AlpeshKumbhare
AlpeshKumbhare

Posted on

Building Agentic AI on AWS: From Bedrock Agents to Multi-Agent Orchestration with AgentCore

2026 is the year AI moved from "answer questions" to "take actions." Agentic AI — systems that autonomously plan, reason, use tools, and execute multi-step tasks — has become the dominant pattern for building intelligent applications on AWS.

This post covers the full agentic AI stack on AWS: from single-agent basics to multi-agent orchestration, the infrastructure that runs them, and the guardrails that keep them safe in production.

What Makes AI "Agentic"?

Traditional AI: User asks question → Model generates answer → Done.

Agentic AI: User states goal → Agent plans steps → Agent calls tools → Agent evaluates results → Agent iterates → Goal achieved.

The difference is autonomy. An agent decides what to do, executes actions, and self-corrects — without human intervention at each step.

┌─────────────────────────────────────────────────────────────┐
│                    AGENTIC AI LOOP                            │
│                                                              │
│   User Goal → Plan → Act → Observe → Reason → Act → Done   │
│                  ↑                               │           │
│                  └───────── iterate ──────────────┘           │
└─────────────────────────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode

The AWS Agentic AI Stack

┌─────────────────────────────────────────────────────────────────┐
│  APPLICATION LAYER                                               │
│  Amazon Q (Business & Developer) | Custom agents via Bedrock    │
├─────────────────────────────────────────────────────────────────┤
│  AGENT FRAMEWORKS                                                │
│  Bedrock Agents | Strands Agents SDK | LangGraph on AgentCore   │
├─────────────────────────────────────────────────────────────────┤
│  AGENT INFRASTRUCTURE (AgentCore)                                │
│  Runtime | Memory | Identity | Observability | Code Interpreter │
├─────────────────────────────────────────────────────────────────┤
│  TOOLS & KNOWLEDGE                                               │
│  AgentCore Gateway (MCP) | Knowledge Bases (RAG) | Action Groups│
├─────────────────────────────────────────────────────────────────┤
│  SAFETY & GOVERNANCE                                             │
│  Guardrails | IAM | CloudTrail | Model Evaluation               │
├─────────────────────────────────────────────────────────────────┤
│  FOUNDATION MODELS                                               │
│  Claude | Nova | Llama | Mistral | DeepSeek (via Bedrock)       │
└─────────────────────────────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode

Amazon Bedrock Agents: The Managed Path

Bedrock Agents is the fully managed way to build AI agents. You define the agent's instructions, connect tools and knowledge, and Bedrock handles the orchestration loop (ReAct-style reasoning).

Core Concepts

Concept What It Does
Instructions System prompt that defines agent's role, behavior, and boundaries
Action Groups Tools the agent can call (Lambda functions, APIs, or return-of-control)
Knowledge Bases RAG — grounds agent responses in your data (documents, databases)
Guardrails Safety controls (content filters, PII masking, denied topics)
Memory Session persistence — agent remembers context across turns
Code Interpreter Agent can write and execute code to solve problems

How the Orchestration Loop Works

  1. User sends a message to the agent
  2. Agent's foundation model reasons about what to do (using instructions + context)
  3. Agent decides to call a tool (action group) or query knowledge (RAG)
  4. Tool executes and returns results
  5. Agent evaluates the results — decides if goal is met or needs more steps
  6. Repeat until goal is achieved or max iterations reached
  7. Agent returns final response to user

Building an Agent: Key Design Decisions

Choosing the model: Claude Sonnet or Nova Pro for complex reasoning. Haiku or Nova Micro for simple routing agents.

Instruction design: Be specific about the agent's role, what it should NOT do, and how to handle ambiguity. Vague instructions lead to unpredictable behavior.

Tool design: Each tool should do ONE thing well. Name them clearly (the model uses the name and description to decide when to call them). Include input/output schemas.

# Example: Defining an action group tool
{
  "actionGroupName": "OrderManagement",
  "description": "Manages customer orders - lookup, modify, cancel",
  "apiSchema": {
    "payload": "openapi-schema.json"
  },
  "actionGroupExecutor": {
    "lambda": "arn:aws:lambda:us-east-1:123456789:function:order-api"
  }
}
Enter fullscreen mode Exit fullscreen mode

Multi-Agent Collaboration: Supervisor Pattern

For complex problems, a single agent isn't enough. Multi-agent collaboration lets specialized agents work together:

Architecture: Supervisor + Collaborators

                    ┌──────────────────┐
                    │  Supervisor Agent │
     User ────────→│  (Routes tasks)  │
                    └────────┬─────────┘
                             │
              ┌──────────────┼──────────────┐
              │              │              │
              ▼              ▼              ▼
     ┌──────────────┐ ┌───────────┐ ┌──────────────┐
     │ Research Agent│ │ Code Agent│ │ Review Agent │
     │ (RAG + Web)  │ │ (CodeGen) │ │ (Validation) │
     └──────────────┘ └───────────┘ └──────────────┘
Enter fullscreen mode Exit fullscreen mode

How It Works on Bedrock

  1. Supervisor agent — Receives user request, analyzes complexity, and routes to specialist agents
  2. Collaborator agents — Each has specific tools and knowledge. Execute their specialty and return results to supervisor.
  3. Supervisor synthesizes — Combines collaborator outputs into final response

When to Use Multi-Agent

Scenario Single Agent Multi-Agent
FAQ chatbot Overkill
Code generation only Unnecessary
Research + summarize + format ⚠️ Gets messy ✅ Clean separation
Customer support (billing + tech + shipping) ⚠️ Tool overload ✅ Specialist routing
Complex analysis with validation ⚠️ Context window limits ✅ Divide and conquer

Rule of thumb: If one agent would need >10 tools or >3 distinct responsibilities, split into multiple agents.


Amazon Bedrock AgentCore: Production Infrastructure

AgentCore is the runtime infrastructure for deploying agents at scale. It provides the "boring but critical" capabilities agents need in production:

AgentCore Components

Component Purpose
Runtime Serverless execution environment for agents (auto-scaling, isolation)
Memory Managed long-term memory across sessions (agent remembers past interactions)
Identity Authentication for agent-to-service and agent-to-agent communication
Observability Traces, metrics, and logs for debugging agent behavior
Code Interpreter Sandboxed code execution (Python/JS) for data analysis tasks
Gateway Converts APIs and Lambda functions into MCP-compatible tools

AgentCore Gateway: The Tool Layer

The Gateway is particularly powerful — it transforms your existing APIs into tools that any agent can discover and use via the Model Context Protocol (MCP):

  • Point Gateway at your OpenAPI spec or Lambda function
  • Gateway auto-generates MCP-compatible tool definitions
  • Agents discover tools at runtime (no hard-coding)
  • Security: IAM-based access control per tool

This means agents don't need tools baked into their code. They discover capabilities dynamically — add a new API to Gateway, and all connected agents can immediately use it.


Strands Agents SDK: The Open-Source Path

For teams wanting more control, AWS released Strands Agents SDK — an open-source Python framework for building agents that runs on AgentCore:

Why Strands?

  • Model-agnostic — works with any Bedrock model (or external models)
  • Tool-first — tools are defined with @tool decorator, auto-generating schemas
  • Memory built-in — integrates with AgentCore Memory for persistence
  • MCP native — discovers tools from MCP servers at runtime
  • Observable — built-in tracing compatible with AgentCore Observability

Basic Agent Structure

from strands import Agent, tool
from strands.models.bedrock import BedrockModel

@tool
def get_weather(city: str) -> str:
    """Get current weather for a city."""
    # Call weather API
    return f"Weather in {city}: 22°C, sunny"

@tool
def create_ticket(title: str, priority: str) -> str:
    """Create a support ticket in the ticketing system."""
    # Call ticketing API
    return f"Created ticket: {title} (priority: {priority})"

agent = Agent(
    model=BedrockModel(model_id="anthropic.claude-sonnet-4-20250514"),
    tools=[get_weather, create_ticket],
    system_prompt="You are a helpful assistant that can check weather and create tickets."
)

response = agent("Check the weather in London and create a ticket if it's raining")
Enter fullscreen mode Exit fullscreen mode

Strands vs Bedrock Agents: When to Use Which

Criteria Bedrock Agents (Managed) Strands SDK (Code-first)
Setup complexity Low (console/API) Medium (write code)
Customization Moderate Full control
Orchestration logic AWS-managed ReAct loop Custom (you define the loop)
Multi-agent Built-in supervisor pattern Build your own topology
Deployment Fully managed AgentCore Runtime or self-hosted
Best for Standard use cases, rapid prototyping Complex custom logic, advanced patterns

Knowledge Bases: Grounding Agents in Facts

Without knowledge, agents hallucinate. Knowledge Bases provide RAG (Retrieval-Augmented Generation):

How It Works

  1. Ingest — Upload documents (PDF, HTML, Markdown, Word, CSV) to S3
  2. Chunk & Embed — Knowledge Base splits documents into chunks, generates embeddings
  3. Store — Embeddings stored in vector database (OpenSearch Serverless, Aurora, Pinecone, or Managed KB)
  4. Retrieve — When agent needs information, relevant chunks are retrieved and injected into prompt
  5. Generate — Model generates response grounded in retrieved facts

Managed Knowledge Base (GA June 2026)

The latest option — fully managed RAG without provisioning anything:

  • No vector database to manage
  • Auto-scaling retrieval
  • Multimodal ingestion (text, images, tables)
  • Built-in re-ranking for relevance
  • Agentic retrieval (multi-hop reasoning across documents)

Guardrails: Keeping Agents Safe

Agents that take actions need safety boundaries. Bedrock Guardrails provides:

Policy Type What It Does
Content filters Block harmful content (hate, violence, sexual, misconduct) with configurable thresholds
Denied topics Prevent agent from discussing specific topics (competitor info, legal advice, etc.)
Word filters Block specific words or phrases
Sensitive information Detect and mask PII (names, emails, credit cards, SSNs)
Grounding check Detect hallucinations by comparing response against source documents
Contextual grounding Verify response relevance to the user's query

Applying Guardrails

Guardrails attach to:

  • The agent itself (all interactions filtered)
  • Specific Knowledge Base queries
  • Individual nodes in a Bedrock Flow

Key insight: Apply guardrails on BOTH input (what users send) AND output (what agents respond). Users can craft prompts to bypass instructions — guardrails are the defense layer.


Orchestration Patterns

Pattern 1: Supervisor-Worker (Hierarchical)

Best for: Customer support, multi-domain queries.
One supervisor routes to specialist workers. Workers don't talk to each other.

Pattern 2: Pipeline (Sequential)

Best for: Document processing, content creation.
Agent A → Agent B → Agent C. Each stage enriches output.

Pattern 3: Parallel Fan-Out

Best for: Research, data gathering from multiple sources.
Multiple agents work simultaneously, results aggregated.

Pattern 4: Debate/Validation

Best for: High-stakes decisions, code review.
Generator agent produces output, critic agent evaluates quality, iterate until criteria met.


Production Checklist for Agentic AI

Before deploying agents to production:

  • [ ] Guardrails configured — content filters, denied topics, PII masking
  • [ ] IAM scoped — agent's execution role has minimum required permissions
  • [ ] Tool permissions bounded — each tool can only access specific resources
  • [ ] Observability enabled — traces for every agent invocation (debug failed reasoning)
  • [ ] Cost controls — max iterations per invocation, token budgets
  • [ ] Fallback behavior defined — what happens when agent can't solve the problem?
  • [ ] Human-in-the-loop — for high-impact actions (delete, purchase, deploy), require approval
  • [ ] Testing — evaluate against known good/bad inputs before production
  • [ ] Rate limiting — prevent runaway agents from flooding APIs
  • [ ] Audit trail — CloudTrail logging of all agent actions and tool invocations

What's Coming Next

The agentic AI space on AWS is evolving rapidly:

  • Bedrock Flows — visual builder for chaining agents, prompts, and conditions without code
  • Agent-to-agent communication — agents that discover and delegate to other agents autonomously
  • Long-running agents — agents that persist across hours/days (not just request-response)
  • AgentCore Memory improvements — structured memory with entity relationships, not just conversation history
  • MCP ecosystem growth — more pre-built tool servers for common services (databases, APIs, SaaS platforms)

Summary

Building agentic AI on AWS in 2026:

  1. Start with Bedrock Agents for managed orchestration — fast to prototype, production-ready
  2. Use Strands SDK when you need custom orchestration logic or advanced patterns
  3. Deploy on AgentCore for production infrastructure (memory, identity, observability)
  4. Connect tools via AgentCore Gateway — MCP-based discovery, zero hard-coding
  5. Ground with Knowledge Bases — RAG prevents hallucination
  6. Protect with Guardrails — content filters, PII masking, grounding checks on every interaction
  7. Scale with multi-agent patterns — supervisor-worker for complex domains, pipeline for sequential processing

The shift from "chatbot that answers" to "agent that acts" is the defining pattern of cloud AI in 2026. The infrastructure is ready — the question is what you build on it.


Alpesh Kumbhare is an AWS Architect at Atos, specializing in AWS infrastructure automation and cloud AI solutions. Connect on LinkedIn.

Top comments (0)