Skip to content
Agentic Control Plane

ACP and Bedrock AgentCore: how the two layers compose

David Crowe David Crowe · · 4 min read
aws bedrock agentcore multi-cloud governance comparison
Share X HN LinkedIn

Run agents? ACP lets you see, control, and price every tool call they make — allow/ask/deny policy, per-session cost, and a full audit log for Claude Code, Cursor, Codex, and OpenClaw, in one command:

macOS · Linux · WSL
curl -sf https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

Getting started →  ·  see your first governed call →  ·  free up to 5 agents

Do you need ACP if you’re on AWS? If every agent you run lives inside Bedrock, probably not. AgentCore is the right main tool. If any of them run somewhere else, you need a second layer.

AWS shipped Bedrock AgentCore Policy as generally available in March 2026. AgentCore Identity, Gateway and Registry (in preview) sit beside it. The result overlaps a lot with what we build, and AWS describes it in similar words: a trust layer for enterprise AI. Everything below about AgentCore is as of April 2026.

What AgentCore covers

  • AgentCore Policy. You write policies in natural language and they compile to Cedar. They’re enforced at AgentCore Gateway, which checks the call before the model can reason around it.
  • AgentCore Identity. Carries the user’s identity across AWS service boundaries, with JWT/OIDC trusted issuers and scoped credentials for downstream calls.
  • AgentCore Gateway. Intercepts tool calls and applies access controls based on the user and the tool’s input parameters.
  • AgentCore Registry (preview). A central catalog of agents and tools.
  • CloudWatch logs. Every enforcement decision is recorded.

For a team running Claude or OpenAI models on Bedrock, this is the real thing. Policies sit outside agent code, new agents inherit the rules, and the security team can run the policy engine without touching the agents.

If your agents live in Bedrock, use it.

What’s outside AgentCore

AgentCore checks calls at AWS’s boundary. Three kinds of agent activity are outside that boundary by design.

1. Coding agents on developer laptops. Claude Code, Cursor, Codex CLI and Zed run on the developer’s machine and talk to the model provider directly. They don’t go through Bedrock. The place to check them is the agent’s own hooks, which run before and after each tool call.

2. Agent code that runs anywhere else. A CrewAI app on Render. A LangGraph supervisor on Cloud Run. An Anthropic Agent SDK app on Fly.io. A Pydantic AI agent on your own Kubernetes. None of these are hosted by AWS, so AgentCore doesn’t see them. The check has to sit in the agent code, around each tool.

3. More than one cloud. AgentCore unifies governance inside AWS, which is its job. One audit log across AWS, GCP and Azure is a different job.

These aren’t flaws in AgentCore. They’re the edge of what an AWS control plane is built to cover.

What ACP adds

  • One audit log across environments. AWS-hosted agents, GCP-hosted agents, Azure-hosted agents and laptop coding agents land in the same log, so “what did our AI do for user X this week?” is one query. How much each source records depends on the integration. The coverage page lists what each coding agent supports.
  • More frameworks and clients. There are more than 30 integration guides. They include Claude Code, Cursor, Codex CLI, Zed, Cline, Claude Desktop, the Anthropic Agent SDK, the OpenAI Agents SDK, Google ADK, CrewAI, LangGraph, Mastra, Pydantic AI, the Vercel AI SDK and AutoGen. AgentCore covers one shape: agents hosted on Bedrock.
  • Identity from your own identity provider. AgentCore Identity works well if you use Cognito or can trust an external OIDC issuer for AWS calls. ACP takes the end user’s token from your provider (Auth0, Okta, Entra, Firebase and others) and carries it through to APIs that aren’t on AWS.

Three ways to combine them

Pattern 1: AgentCore for AWS, ACP for the rest

This is the usual setup for multi-cloud teams or teams with many developer tools:

  • Bedrock agents use AgentCore Policy and Gateway.
  • Agents elsewhere (CrewAI on Render, LangGraph on Cloud Run, Cursor on laptops) use ACP.
  • For one view across both, send CloudWatch logs to your SIEM next to ACP’s audit log, which you can export as CSV, and join on session or request IDs.

Pattern 2: ACP as an MCP server for AgentCore agents

If your AgentCore agents can call MCP servers, add ACP at https://api.agenticcontrolplane.com/mcp. The agent can then call acp_check to ask whether a step would be allowed, and acp_status to read your workspace mode.

That is advisory. ACP’s MCP endpoint serves only its own governance tools. It doesn’t run your action groups, and it doesn’t scan their output. For enforcement and output scanning, use Pattern 3 or a hook.

Pattern 3: ACP inside the tool handler

Say your AgentCore agent has an Action Group backed by a Lambda that calls a non-AWS endpoint, such as your CRM. AgentCore Gateway can allow or deny the call, but it can’t see what the tool does after that, including PII in the response.

Wrap the Lambda handler with @governed. The response then gets PII scanning and secret redaction, and an audit row is written before it goes back to the agent. AgentCore decides who can call the tool. ACP checks what comes back.

When AgentCore alone is enough

AgentCore alone works if all of these are true:

  • All your agents run inside Bedrock.
  • You have no Claude Code, Cursor or other coding agents to cover.
  • You’re on AWS only.
  • Your identity provider is Cognito, or you’re fine mapping every outside provider into AWS identity.

If that’s you, skip ACP. Otherwise the question isn’t which one to pick. It’s how to run them together.

Install steps are at /integrations/aws-bedrock.

Reference architecture → · Get started →

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
Related posts

← back to blog