Amazon Bedrock AgentCore: The Current Agent Platform — Agents for Bedrock & AgentCore

Agents Are Workflow Systems

Amazon Bedrock Agents (Classic), launched in November 2023, is now called Bedrock Agents Classic: it entered maintenance mode (closed to new customers on July 30, 2026), succeeded by Amazon Bedrock AgentCore — AWS's modular, framework-agnostic agent platform that works with any foundation model and any agent SDK (Strands Agents, LangGraph, CrewAI, LlamaIndex, Google ADK, OpenAI Agents SDK).

AgentCore's core services can be used independently or together:

• Runtime — serverless, session-isolated microVM hosting (up to 8-hour workloads) for agents and tools with sub-second cold starts via BedrockAgentCoreApp.
• Gateway & Policy — turns OpenAPI/REST endpoints, AWS Lambda functions, and Knowledge Bases into Model Context Protocol (MCP) tools with semantic tool discovery and deterministic Cedar policy guardrails.
• Memory — short-term (multi-turn session events) and long-term (cross-session extracted facts & user preferences) memory, shareable across agent swarms.
• Identity & Consent Portal — workload identity + secure OAuth/API-key token vault that plugs into Cognito, Okta, and Entra ID, featuring the AgentCore Consent Portal for granular, time-bound SaaS delegation.
• Code Interpreter & Browser — sandboxed Python/JS/TS code execution (with S3 file IO) and managed cloud browser runtime driven via Chrome DevTools Protocol (CDP).
• AWS Agent Registry — centralized enterprise catalog to discover, version, test, and enforce IAM boundaries across company-wide agents and MCP servers.
• Observability & Evaluations — distributed OpenTelemetry / W3C traceparent spans in CloudWatch GenAI dashboards plus LLM-as-a-judge evaluation pipelines.

Strands Agents SDK + Bedrock AgentCore Runtime (Python)

AWS's open-source Strands Agents SDK (pip install strands-agents bedrock-agentcore) pairs model-driven agent loops with native MCP tool discovery and one-decorator AgentCore deployment:

from bedrock_agentcore import BedrockAgentCoreApp
from strands import Agent
from strands.models import BedrockModel
from strands.tools.mcp import MCPClient

app = BedrockAgentCoreApp()
model = BedrockModel(model_id="us.anthropic.claude-sonnet-5-5", region_name="us-east-1")
mcp_gateway = MCPClient(server_command="uvx", server_args=["aws-docs-mcp-server"])

agent = Agent(
    model=model,
    tools=mcp_gateway.get_tools(),
    system_prompt="You are a production cloud SRE agent with bounded MCP tools.",
)

@app.entrypoint
async def invoke(payload):
    async for event in agent.stream_async(payload["prompt"]):
        yield event