Amazon Bedrock Agents launched in November 2023 as AWS's managed way to build agents on Bedrock-hosted models. As of 2026 it has been renamed Amazon Bedrock Agents Classic, and per AWS's documentation it "will no longer be open to new customers starting July 30, 2026." Existing customers can continue using the service as normal — this is a soft-close, not a shutdown — but AWS explicitly points new users to Amazon Bedrock AgentCore for similar capabilities.
What Classic was, and who is affected
Bedrock Agents Classic supports all models supported by Amazon Bedrock, so Claude models work there without any Claude-specific allowlist. The service runs in 16 regions, including us-east-1, us-west-2, eu-central-1, ap-northeast-1, and us-gov-west-1. If your organization already runs Classic agents in production, nothing forces you to move today — but a service closed to new customers is a service whose long-term roadmap deserves skepticism, and it is reasonable to treat new feature investment there as unlikely. The practical question is what to build on instead.
Three candidate paths for Claude agents on AWS
Teams standardizing on Claude have three realistic options after Classic, and they differ mainly in who runs the agent loop — the repeated cycle of model reasoning, tool calls, and result handling.
| Bedrock AgentCore | Claude Agent SDK | Claude Managed Agents | |
|---|---|---|---|
| Operated by | AWS | You (library in your process) | Anthropic |
| Model scope | Any foundation model, incl. Claude | Claude only | Claude only |
| Runs against Bedrock? | Yes (native) | Yes (CLAUDE_CODE_USE_BEDROCK=1) | No — 1P and Claude Platform on AWS only (beta) |
| Execution environment | Isolated microVM per session | Your infrastructure | Anthropic-managed container per session |
Amazon Bedrock AgentCore is AWS's designated successor: a family of services (Harness, Runtime, Memory, Gateway, Identity, and more) that work together or independently, with the Harness running a managed agent loop in an isolated microVM per session. It is model-agnostic — Claude is one supported model among OpenAI, Gemini, Nova, Llama, and Mistral — and its "skills" and MCP support are AWS-native concepts, not Anthropic's same-named features. See the AgentCore architecture overview for the full picture. Choose this when you want AWS to own the runtime and may mix model vendors.
The Claude Agent SDK is Anthropic's documented agent path on Bedrock. It packages the same tools, agent loop, and context management that power Claude Code, as a library (claude-agent-sdk on pip, @anthropic-ai/claude-agent-sdk on npm) that runs in your process on your infrastructure. Set CLAUDE_CODE_USE_BEDROCK=1 and it authenticates through the standard AWS SDK credential chain and bills through your Bedrock account. Know the documented Bedrock limitations: the built-in WebSearch tool is unavailable there, and the SDK uses the Bedrock Invoke API rather than Converse. Details in running the Agent SDK on Bedrock. Choose this when you want Anthropic's agent behaviors — hooks, subagents, filesystem skills, local MCP servers — with Bedrock's billing and IAM.
Claude Managed Agents is Anthropic's hosted agent harness: a REST API where Anthropic runs the loop and a sandboxed container per session. It is the least-infrastructure option, but the availability constraint is decisive here: Managed Agents is in beta on the first-party Claude API and Claude Platform on AWS only — it is not available on Amazon Bedrock. If your requirement is "everything through our Bedrock account," it is off the table; if "runs on AWS" is enough, Claude Platform on AWS (Anthropic-operated, on AWS infrastructure) qualifies. Compare in Agent SDK vs. Managed Agents.
What to evaluate before committing
Whichever way you lean, test three things with a pilot before migrating anything off Classic: identity and credential flow (IAM roles for AgentCore and the Agent SDK; workspace keys for Managed Agents), tool availability on your chosen surface (server-side tools vary sharply by platform — check the feature matrix), and observability (where do traces, token counts, and errors land, and can your existing monitoring see them). Migration effort is dominated by tool re-wiring, not by prompts.