---
title: How do you govern AI agents built on Amazon Bedrock AgentCore?
url: https://opencontroller.lyzr.com/blog/govern-amazon-bedrock-agentcore-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern AI agents built on Amazon Bedrock AgentCore?

**Govern AgentCore agents in two layers. Inside AgentCore, use Identity for inbound and outbound auth, route every tool call through Gateway so Cedar-based Policy can enforce default-deny rules, and turn on Observability and Evaluations. Then register each agent in an estate-wide inventory, because AgentCore's controls stop at the agents and tool calls that run through it.**

AgentCore is a set of services, not one product. AWS says each service can be used on its own, with any framework and any model. That flexibility helps builders. It also means governance depends on which services a team actually switched on. This guide maps each service to a governance question and shows where the gaps are.

## Which AgentCore services matter for governance?

AgentCore became generally available in October 2025. Policy reached general availability on 3 March 2026, and Evaluations followed in March 2026.

| Governance question | AgentCore service | What it gives you |
|---|---|---|
| Who can call this agent or tool? | Gateway inbound auth | IAM SigV4, a custom JWT/OAuth authorizer, authenticate-only, or none |
| Which credentials does the agent use downstream? | Identity token vault | OAuth tokens and API keys encrypted with KMS; 2-legged and 3-legged OAuth |
| Which tools can it call, with which inputs? | Gateway + Policy | Cedar policies checked on every tool call through the Gateway; default-deny, forbid wins |
| Is it isolated from other users? | Runtime | Each session runs in its own microVM; memory is wiped when the session ends |
| What did it do? | Observability, CloudTrail | OpenTelemetry-compatible telemetry in CloudWatch; CloudTrail events for Gateway and Agent Registry |
| Is it good enough to ship? | Evaluations | 13 built-in evaluators, including Goal Success Rate, Tool Selection Accuracy and Harmfulness |

## How should you set up identity and access?

Start with inbound auth on every Gateway. The [Gateway inbound auth docs](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-inbound-auth.html) list four modes and warn against using no authentication in production. When agents act for people, use a JWT authorizer backed by your identity provider. JWT claims are then recorded in CloudTrail, which gives you the "who asked" half of every audit question.

For outbound calls, use credential providers instead of secrets in code. They can use the gateway's IAM role, an OAuth client-credentials flow or an API key. The token vault encrypts what it stores with KMS, using an AWS-managed or a customer-managed key. Use 3-legged OAuth when an agent needs a user's consent to reach a SaaS system, and 2-legged OAuth for machine-to-machine calls. Our guide to [on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation) explains when each pattern fits.

## How does AgentCore Policy enforce tool rules?

A policy engine is attached to a Gateway and evaluates every tool call that passes through it. Policies are written in Cedar with default-deny and forbid-wins semantics, so nothing runs unless a rule permits it, and any matching forbid overrides a permit. You can draft rules in plain English, and automated reasoning flags rules that look too permissive or too restrictive. Bedrock Guardrails scores can also feed policy decisions.

The limit is the traffic path. Policy only sees tool calls that go through a Gateway. If an agent calls an API directly from its own code, no Cedar rule is checked. Make "all tools behind a Gateway" a deployment rule, and check agent egress to catch exceptions.

## What should an AgentCore governance checklist include?

1. Every Runtime agent has a named owner and appears in your inventory.
2. No Gateway uses the no-auth inbound mode.
3. Outbound credentials live in the token vault, encrypted with a customer-managed KMS key if your policy requires one.
4. Every tool is exposed through a Gateway; direct egress is blocked or reviewed.
5. A policy engine is attached to each Gateway, and its Cedar rules are reviewed like code.
6. CloudWatch Transaction Search is on, so OpenTelemetry traces are captured, and retention matches your [log retention policy](/blog/how-long-to-retain-ai-agent-logs).
7. Evaluations run before each release and on samples of production traffic.
8. Session limits are set on purpose. A Runtime session can last up to 8 hours per lifecycle, with a 15-minute idle default.
9. A CloudTrail trail captures Gateway events and is retained centrally.

## What does AgentCore not cover?

AgentCore governs what runs on AgentCore. Four gaps remain for most enterprises:

- **Other platforms.** Agents on Azure, Google Cloud, SaaS platforms and developer laptops are outside its view.
- **Bypassed tool calls.** Calls that skip the Gateway skip Policy.
- **Partial audit.** The docs name CloudTrail coverage for Gateway and Agent Registry, not every capability.
- **Estate-wide questions.** Total model spend per agent across providers, or every agent that shares an over-broad role, need data from more than one platform.

Closing those gaps is the job of an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane). It does not replace AgentCore. It sits above it and above every other platform you run.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and has a runtime adapter for Bedrock AgentCore. AgentCore agents land in the same governed registry as agents from other platforms, with owners, versions, tools and MCP servers. Posture flags risky permissions and configuration. Rogue-agent handling watches for behavioural deviation and can quarantine or kill an agent. Policies are code, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO/IEC 42001. Registration is not runtime takeover: how deep enforcement goes depends on the connector and the traffic path. To see it on your own AWS estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does AgentCore Policy apply to every tool call an agent makes?**
No. Policy evaluates tool calls that pass through an AgentCore Gateway. Calls an agent makes directly from its code are not checked.

**What policy language does AgentCore Policy use?**
Cedar, with default-deny and forbid-wins semantics. You can also draft rules in natural language, and automated reasoning flags rules that are too permissive or too restrictive.

**Can AgentCore govern agents that run outside AWS?**
AgentCore services can be used with any framework or model, and telemetry can be sent through the AWS Distro for OpenTelemetry. Its policy and runtime controls still apply only to traffic that runs through AgentCore, so agents on other clouds and SaaS platforms need a separate inventory and control layer.

## Key facts

Amazon Bedrock AgentCore went GA in October 2025; AgentCore Policy went GA on 3 March 2026. Policy uses Cedar, attaches to a Gateway, is default-deny with forbid-wins, and only evaluates tool calls routed through the Gateway. Runtime sessions run in isolated microVMs for up to 8 hours per lifecycle. Opencontroller (by Lyzr) discovers agents across AWS, Azure and Google Cloud and has a runtime adapter for Bedrock AgentCore. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [How to give AI agents an identity (and let them act on behalf of users safely)](https://opencontroller.lyzr.com/blog/ai-agent-identity-on-behalf-of-delegation.md)
- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
