AWS AgentCore Is Here: What to Do Before Classic Locks You Out

AWS closed Bedrock Agents Classic to new customers on July 30, 2026, freezing its model catalog and making AgentCore the only path to newer foundation models. AgentCore is a ground-up re-architecture with five independently adoptable pillars—Runtime, Identity, Memory, Gateway, and Browser—and the Gateway feature converts existing Lambda functions into MCP-compatible tools. AWS provides an Import-Agent CLI to migrate Classic agents to AgentCore, but existing Classic workloads continue to run with no announced end-of-life date.

Amazon closed Bedrock Agents Classic to new customers on July 30, 2026. If your architecture doc still says “Bedrock Agents” for a new project, that path is gone. AgentCore is the replacement — and it is not a rename. This is a ground-up re-architecture with a different control plane and a different mental model. Teams that discovered this by trying to enable AgentCore Memory on a Classic agent alias found out the hard way: they are separate systems. The Clock That Matters: The Model Freeze Existing Classic workloads keep running. AWS has not announced an end-of-life date, and existing users will not be cut off suddenly. That might tempt you to defer the migration. Do not. The forcing function is the model catalog, and it froze on July 30. Every foundation model AWS ships after that date — newer Claude versions, updated Titan variants, new third-party additions to Bedrock — is available only through AgentCore. Classic users are locked to whatever was in the catalog before maintenance mode began. If your product roadmap requires a better model in the next six to twelve months, you need to start planning your migration now, not when AWS eventually announces a hard deprecation. What AgentCore Actually Is Classic was one managed service with a single configuration surface: you declared a model, action groups, a knowledge base, and guardrails, and AWS handled orchestration. Simple, but deeply opinionated. AgentCore takes the opposite approach. AgentCore is five independently adoptable pillars: Runtime — Serverless agent compute with microVM-per-session isolation. No Docker builds, no ECR pushes, no Kubernetes. Scales to thousands of concurrent sessions. Sync timeout is 15 minutes; async sessions run up to 8 hours, making long-running agents finally practical on AWS. Identity — OAuth and JWT handling so your agent can authenticate to third-party services GitHub, Slack, Salesforce without you building the token management layer. Memory — Persistent memory SDK with native LangChain and LangGraph integration. Session memory and cross-session memory handled as a managed service. Gateway — Converts AWS Lambda functions and REST APIs into MCP-compatible tools. No protocol code, no server to host. Browser — Managed browser instances for web automation workloads at agent scale. Each pillar is independently adoptable. Start with Runtime, add Gateway when you need tool connectivity, layer in Identity when your agent starts accessing user-specific data. You do not have to migrate everything at once, and that incremental path is genuinely useful. Gateway Is the Feature You’re Underestimating AgentCore Gateway deserves more attention than it is getting. Point it at an existing Lambda function, give it a tool schema, and that function is immediately available to any MCP-compatible agent as a callable tool — no server, no protocol boilerplate, no rewrite. It now supports the MCP 2026-07-28 stateless spec https://aws.amazon.com/blogs/machine-learning/how-agentcore-gateway-supports-the-mcp-2026-07-28-spec/ — the largest MCP revision since launch. Migrating to the new spec is a single UpdateGateway API call. For teams that have been stalled on “how do we expose our tools to agents,” Gateway is a real answer. Your existing Lambda backend becomes agent-callable with minimal additional work. If there is one AgentCore feature to evaluate before committing to a full migration, this is it. How to Migrate from Bedrock Agents Classic AWS provides an Import-Agent CLI that converts an existing Classic agent to a LangGraph or Strands scaffold and optionally deploys it to AgentCore Runtime: agentcore import-agent --region us-east-1 --agent-id YOUR AGENT ID --agent-alias-id YOUR ALIAS ID --target-platform strands --output-dir ./migrated-agent --deploy-runtime The tool handles agent selection, framework targeting, primitive configuration, and deployment. For full migration documentation, see the Amazon Bedrock Agents Classic maintenance mode guide https://docs.aws.amazon.com/bedrock/latest/userguide/agents-classic-maintenance-mode.html and the AgentCore developer guide https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html . AWS’s two-to-four week estimate is realistic for straightforward agents. It gets longer fast: | Agent Complexity | Migration Effort | Primary Approach | |---|---|---| | Model + action groups + knowledge base | Hours to 1 day | Import-Agent CLI | | Custom orchestration logic | 1–3 weeks | Manual rebuild with Runtime | | Multi-agent, complex graphs | 3–6 weeks | Full framework migration | Community reports from teams migrating non-trivial production systems are running closer to six weeks. Budget accordingly. The Bigger Question: Platform Longevity Bedrock Agents launched in November 2023 and was closed to new customers by July 2026 — a 32-month lifecycle for what was positioned as a production-grade AI service. Teams that built on it were following AWS’s recommended path. But AI infrastructure is moving faster than traditional cloud services, and managed AI agent platforms have not yet demonstrated the kind of long-term stability that enterprise tooling usually requires. AgentCore’s framework-agnostic design is a direct response to this problem. If AWS changes direction again, your Strands or LangGraph code is portable. The AgentCore announcement https://aws.amazon.com/blogs/aws/introducing-amazon-bedrock-agentcore-securely-deploy-and-operate-ai-agents-at-any-scale/ explicitly frames it as infrastructure you own rather than a managed service you’re handed. That’s the right architectural direction. Whether it holds over the next three years is the question no roadmap can answer today. What to Do Right Now For new projects on AWS: start with AgentCore Runtime. Classic is not available to you. For teams running Classic in production: audit your model roadmap. If you need any new foundation model in the next twelve months, that is your migration trigger date — work backwards from it. If you are content with the current model catalog, you have more runway, but the migration only gets harder as Classic ages and AgentCore diverges. Either way, start with Gateway. Convert your existing Lambda tools to MCP, validate the observability and routing improvements, and confirm that AgentCore Runtime fits your workload before committing to a full migration. It is the lowest-friction entry point and it pays off immediately — regardless of where the rest of your migration lands.