AWS Agent Backend Implementations

Status: 🟢 In-use (Part of consolidated AWS integration) Category: Integrations Date: October 27, 2025 (Consolidated into 2025-11-05-aws-integration.md)

The AWS agent supports two backend implementations:

1. LangGraph Backend (Default) ✨

File: agent_aws/agent_langgraph.py

Features:

• ✅ Tool Call Notifications: Shows 🔧 Calling tool: {ToolName} and ✅ Tool {ToolName} completed
• ✅ Token-by-Token Streaming: Fine-grained streaming when ENABLE_STREAMING=true
• ✅ Consistent with Other Agents: Same behavior as ArgoCD, GitHub, Jira agents
• ✅ LangGraph Ecosystem: Full access to LangGraph features

Usage:

# Default - no configuration needed
docker-compose -f docker-compose.dev.yaml up agent-aws-p2p

# Or explicitly set
export AWS_AGENT_BACKEND=langgraph
export ENABLE_STREAMING=true

Example Output:

🔧 Aws: Calling tool: List_Clusters
✅ Aws: Tool List_Clusters completed

Found 3 EKS clusters in us-west-2:
- prod-cluster
- staging-cluster
- dev-cluster

---

2. Strands Backend (Alternative)

File: agent_aws/agent.py

Features:

• ✅ Chunk-Level Streaming: Built-in streaming (always on)
• ✅ Mature: Original implementation, well-tested
• ✅ Simple: Fewer dependencies
• ❌ No Tool Notifications: Tools are called internally (not visible)
• ❌ No Token-Level Streaming: Streams in larger chunks

Usage:

export AWS_AGENT_BACKEND=strands
docker-compose -f docker-compose.dev.yaml up agent-aws-p2p

Example Output:

Found 3 EKS clusters in us-west-2:
- prod-cluster
- staging-cluster
- dev-cluster

---

Comparison Table

Feature	LangGraph (Default)	Strands
Tool Notifications	✅ Yes (🔧, ✅)	❌ No (internal)
Token Streaming	✅ Yes (with ENABLE_STREAMING=true)	⚠️ Chunk-level only
Streaming Control	✅ Via ENABLE_STREAMING	❌ Always on (chunks)
Agent Name in Messages	✅ Yes	❌ No
Consistency	✅ Matches other agents	⚠️ Different format
Maturity	✨ New	✅ Well-tested
Dependencies	LangGraph, LangChain	Strands SDK

---

Environment Variables

AWS Agent Backend Selection

# Choose the backend implementation
AWS_AGENT_BACKEND=langgraph  # default
# or
AWS_AGENT_BACKEND=strands

Streaming Configuration (LangGraph only)

# Enable token-by-token streaming
ENABLE_STREAMING=true  # default for AWS agent

MCP Configuration (Both backends)

# Enable/disable AWS MCP servers
ENABLE_EKS_MCP=true
ENABLE_COST_EXPLORER_MCP=true
ENABLE_IAM_MCP=true
ENABLE_TERRAFORM_MCP=false
ENABLE_AWS_DOCUMENTATION_MCP=false
ENABLE_CLOUDTRAIL_MCP=true
ENABLE_CLOUDWATCH_MCP=true

---

Recommendation

Use LangGraph backend (default) for:

• ✅ Consistent user experience across all agents
• ✅ Better visibility into tool execution
• ✅ Finer-grained streaming control
• ✅ Better integration with Backstage plugin

Use Strands backend only if:

• You need the original implementation for compatibility
• You're debugging issues with the LangGraph implementation
• You prefer a simpler dependency tree

---

Implementation Details

The executor automatically selects the backend in agent_executor.py:

backend = os.getenv("AWS_AGENT_BACKEND", "langgraph").lower()

if backend == "strands":
    # Use Strands SDK implementation
    from ai_platform_engineering.utils.a2a_common.base_strands_agent_executor import BaseStrandsAgentExecutor
    from agent_aws.agent import AWSAgent
    return BaseStrandsAgentExecutor(AWSAgent())
else:
    # Use LangGraph implementation (default)
    from ai_platform_engineering.utils.a2a_common.base_langgraph_agent_executor import BaseLangGraphAgentExecutor
    from agent_aws.agent_langgraph import AWSAgentLangGraph
    return BaseLangGraphAgentExecutor(AWSAgentLangGraph())

---

Testing Both Implementations

Test LangGraph Backend (Default):

curl -X POST http://localhost:8002 \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"id":"test","method":"message/stream","params":{"message":{"role":"user","parts":[{"kind":"text","text":"list EKS clusters"}]}}}'

# Look for tool notifications:
# 🔧 Aws: Calling tool: ...
# ✅ Aws: Tool ... completed

Test Strands Backend:

export AWS_AGENT_BACKEND=strands
# Restart agent
docker-compose -f docker-compose.dev.yaml restart agent-aws-p2p

curl -X POST http://localhost:8002 \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"id":"test","method":"message/stream","params":{"message":{"role":"user","parts":[{"kind":"text","text":"list EKS clusters"}]}}}'

# No tool notifications, just chunked content