Amazon Bedrock Agents, launched in November 2023 and also known as Agents for Amazon Bedrock, is now called Bedrock Agents Classic. Since 30 July 2026 it is closed to new customers: an AWS account without recent Bedrock Agents use can no longer create an agent. Accounts that used it keep it, without new features. For new work AWS points to Amazon Bedrock AgentCore.
What exactly changed?
| Announced | 30 June 2026, in AWS’s list of services entering maintenance |
| Effective | 30 July 2026 |
| New name | Amazon Bedrock Agents Classic. The API namespace (bedrock-agent), SDK clients, CloudFormation resource types and IAM actions don’t change. |
| Who keeps access | Any account with Bedrock Agents activity in the past 12 months. AWS decides this per account, automatically. |
| Who doesn’t | Accounts without that activity. There is no exception process. |
| End of life | None announced. AWS says there is no migration deadline. |
| New features | None. The model list is frozen as of 30 July 2026; newer models go to AgentCore. |
Only two calls are blocked for accounts without prior use: CreateAgent and InvokeInlineAgent. They fail with:
AccessDeniedException: Bedrock Agents is in Maintenance Mode. New agent creation is not available for accounts without prior service usage.Everything else keeps working for existing users: updating, invoking, aliases, action groups and knowledge base links. Bedrock itself, model inference, Knowledge Bases and Guardrails are not affected and still get new models.
The new-account trap: the allowlist is per account. If your company used Bedrock Agents in one AWS account, a new account for a new project is still locked out. Tutorials from before July 2026 that start with “create an agent in the console” fail there.
What is AgentCore?
AgentCore is AWS’s platform for running agents you build with any framework and any model. Where Bedrock Agents was one configured agent, AgentCore is a set of building blocks:
- Harness: a managed agent loop. You declare a model, tools and instructions, and invoke it with one API call. This is the closest replacement for a classic Bedrock agent.
- Runtime: serverless hosting for agents you write yourself, with LangGraph, CrewAI, LlamaIndex, Google ADK, OpenAI’s Agents SDK, Strands or your own code.
- Gateway: turns your APIs and Lambda functions into MCP tools, and connects existing MCP servers.
- Memory, Identity, Observability: short- and long-term memory, sign-in to other services, and traces in OpenTelemetry format.
- Code Interpreter and Browser: sandboxed code execution and a cloud browser for the agent.
- Policy, Evaluations, Registry and more: rules the agent must follow, automated testing, a catalog of agents and tools.
AgentCore runs models inside and outside Bedrock, including OpenAI, Gemini, Claude and Mistral. It has been generally available since October 2025, the Harness since July 2026.
How do I move a Classic agent to AgentCore?
AWS recommends the Harness unless you need to control the agent loop yourself. Install the AgentCore CLI (Node.js 20.12 or newer):
npm install -g @aws/agentcore
agentcore --versionThen recreate the agent and map its settings:
| Classic agent | AgentCore Harness |
|---|---|
| Foundation model | --model-id |
| Instructions | --system-prompt |
| Action groups | Gateway tools |
| Knowledge bases | A Gateway or retrieval tool |
agentcore create --name my-research-agent
agentcore add harness --name my-research-agent --model-id <model> --system-prompt "<your instructions>"
agentcore deploy
agentcore invoke --harness my-research-agent --session-id test-1 "What can you do?"To keep the agent’s logic as code instead, AgentCore can import a classic agent into a Runtime project:
agentcore add runtime --name imported_agent --agent-id <agent-id> --agent-alias-id <alias-id> --framework strandsAWS also publishes a migration skill for coding agents, in its agent toolkit on GitHub (aws/agent-toolkit-for-aws). Your coding agent then walks through inventory, an eligibility check and a plan, and asks you at each step. It never changes the original agent.
If agentcore behaves strangely, an older Python package may be shadowing the new CLI. AWS’s docs say to run pip uninstall bedrock-agentcore-starter-toolkit.
What doesn’t carry over?
AWS lists three gaps: stage-specific prompt overrides have no direct equivalent, multi-agent setups are limited, and custom orchestrators need Runtime, not the Harness. For simple agents, AWS estimates the move in hours.
What does it cost?
Bedrock Agents Classic itself was free; you paid for the model and attached resources. AgentCore is pay per use with no minimum: Runtime by the second of CPU and memory actually used, Gateway per call, Memory per event and record, and so on, with model usage billed separately. AWS has no separate charge for the Harness’s agent loop. Startup credits from AWS Activate cover AgentCore usage too.
Related guides
- GenAI app builders: what changed across all four big clouds in 2026
- AWS Bedrock: use Bedrock models from Claude Code, Codex or OpenCode
- What is MCP?: the tool standard AgentCore Gateway speaks