---
title: What changes for governance when you move from Bedrock Agents to AgentCore?
url: https://opencontroller.lyzr.com/blog/bedrock-agents-vs-agentcore-governance
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What changes for governance when you move from Bedrock Agents to AgentCore?

**Amazon Bedrock Agents is a managed agent: AWS runs the orchestration loop, and you govern a configuration of instructions, action groups, knowledge bases and guardrails. Amazon Bedrock AgentCore is modular infrastructure for agents built in any framework, so governance moves to separate services: Runtime, Identity, Gateway, Policy, Memory, Observability and Evaluations. You get more control points, and more places where controls can be left off.**

Many AWS estates now run both. Governance teams need one way to inventory and control them, even though the control surfaces differ.

## How do the two services differ?

| Area | Bedrock Agents | Bedrock AgentCore |
|---|---|---|
| Orchestration | AWS-managed loop | Your framework (for example Strands, LangGraph, CrewAI, LlamaIndex) |
| Models | Models available in Bedrock | Bedrock models or external models over API |
| Tools | Action groups backed by Lambda or API schemas | AgentCore Gateway exposes APIs and tools over MCP |
| Knowledge | Bedrock Knowledge Bases | Your choice, plus AgentCore Memory |
| Content safety | Bedrock Guardrails attached to the agent | Bedrock Guardrails if you call them, plus AgentCore Policy |
| Policy outside the model | Limited to configuration and IAM | AgentCore Policy, written in Cedar, evaluated on tool calls |
| Identity | IAM role for the agent | AgentCore Identity for workload identity and acting on behalf of users |
| Execution | Managed | AgentCore Runtime: per-session microVM isolation, sessions up to 8 hours |
| Observability | Agent traces | OpenTelemetry traces into CloudWatch |
| Evaluation | Build your own | AgentCore Evaluations (generally available since 31 March 2026) |

## What gets easier with AgentCore?

- **Policy outside the prompt.** AgentCore Policy evaluates Cedar rules on tool calls through the Gateway, so a rule like "refunds above a limit are denied" does not depend on the model following instructions. See our explainer on [AgentCore Policy](/blog/amazon-bedrock-agentcore-policy-explained).
- **User-scoped access.** AgentCore Identity lets an agent call tools on behalf of a specific user with scoped tokens, instead of one broad role for every request.
- **Session isolation.** Each Runtime session runs in its own microVM, which limits cross-session data leakage. Details in [AgentCore Runtime session isolation](/blog/agentcore-runtime-session-isolation).
- **Standard telemetry.** OpenTelemetry traces make it easier to export evidence to your own tools.

## What gets harder?

- **Controls are opt-in.** A team can deploy to Runtime without routing tools through Gateway, without Policy, and without Evaluations. Bedrock Agents bundled more of this by default.
- **Framework sprawl.** Each team may bring a different framework, prompt format and tool library. Your inventory must record framework and version per agent.
- **More credentials.** External models and third-party tools mean API keys and OAuth clients outside IAM.
- **Model choice outside Bedrock.** Bedrock model access controls and SCPs do not cover calls to external providers made from inside an AgentCore agent.

## What should a governance checklist require for AgentCore agents?

1. Register every Runtime agent with owner, framework, models and tools before production.
2. Route tool access through AgentCore Gateway, or a governed gateway, rather than direct calls.
3. Turn on Policy for any agent with write tools, and keep Cedar policies in version control.
4. Use AgentCore Identity for user-facing agents so actions carry the user's scope.
5. Enable Observability and keep traces for your retention period.
6. Configure online Evaluations with a sampling rate you can afford, plus on-demand evaluations in CI.
7. Route external model calls through an LLM gateway so spend and data paths are visible.

Keep Bedrock Agents on the same list: owner, action groups, knowledge bases, guardrail versions and the IAM role. Our guide to [governing Bedrock AgentCore agents](/blog/govern-amazon-bedrock-agentcore-agents) covers the AgentCore side in depth, and [AgentCore Gateway](/blog/what-is-amazon-bedrock-agentcore-gateway) explains the tool path.

## Should you migrate Bedrock Agents to AgentCore?

Migrate when you need a framework, model or tool pattern that Bedrock Agents cannot express, or when you need policy enforced outside the model. Stay when a managed configuration covers the use case and your team values fewer moving parts. Governance should not force the choice. It should cover both with the same registry, the same policies and the same evidence format.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and puts Bedrock Agents and AgentCore agents in one governed registry with versions, tools and MCP servers. Its runtime adapter for Bedrock AgentCore extends observation and enforcement to AgentCore workloads, and per-agent Block and Quarantine controls act through the connected platform. Registration is not runtime takeover: enforcement depth depends on the connector and traffic path. See the [AI agent control plane](/blog/what-is-an-ai-agent-control-plane) model or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does AgentCore replace Bedrock Agents?**
AWS offers both. Bedrock Agents is the managed option; AgentCore is modular infrastructure for custom frameworks.

**What policy language does AgentCore Policy use?**
Cedar.

**How long can an AgentCore Runtime session run?**
Up to 8 hours, each in an isolated microVM.

## Key facts

Bedrock Agents: AWS-managed orchestration, Lambda action groups, Knowledge Bases, Bedrock Guardrails, Bedrock models. Bedrock AgentCore: framework-agnostic (Strands, LangGraph, CrewAI, LlamaIndex), any model; components Runtime (per-session microVM, up to 8 hours), Identity, Gateway (MCP), Memory, Policy (Cedar), Observability (OpenTelemetry to CloudWatch), Evaluations (GA 31 March 2026), Code Interpreter, Browser. Opencontroller (Lyzr) discovers agents across AWS, Azure, Google Cloud, has a runtime adapter for Bedrock AgentCore and per-agent Block/Quarantine through the connected platform. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern AI agents built on Amazon Bedrock AgentCore?](https://opencontroller.lyzr.com/blog/govern-amazon-bedrock-agentcore-agents.md)
- [How does Policy in Amazon Bedrock AgentCore control agent tool calls?](https://opencontroller.lyzr.com/blog/amazon-bedrock-agentcore-policy-explained.md)
- [How does Amazon Bedrock AgentCore Runtime isolate agent sessions?](https://opencontroller.lyzr.com/blog/agentcore-runtime-session-isolation.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
