Claude Managed Agents vs AWS Bedrock AgentCore

Claude Managed Agents and AWS Bedrock AgentCore both run the agent loop for you, give each session an isolated environment, and let you skip building infrastructure. Teams compare them when deciding whether to build on the model maker's harness or on their cloud provider's.

Claude Managed Agents vs AgentCore at a glance

Claude Managed Agents is Anthropic's pre-built, configurable agent harness, in public beta since April 8, 2026. You define an agent (Claude model, system prompt, tools, MCP servers, skills), an environment, and sessions. Claude runs bash, edits files, searches the web, and calls MCP tools in an Anthropic-managed sandbox or a self-hosted one. Memory stores, credential vaults, webhooks, and scheduled runs are built in, and May 2026 added outcomes and multiagent orchestration in public beta and dreaming in research preview.

Amazon Bedrock AgentCore is AWS's agentic platform for building and operating agents with any framework and any foundation model. It is a set of modular services you can use together or alone: Harness (a managed agent loop, generally available since June 17, 2026), Runtime, Memory, Gateway, Identity, Code Interpreter, Browser, Observability, Evaluations, Optimization, Policy, Registry, and Payments.

Harness and building experience

Claude Managed Agents. One opinionated harness, tuned by the model maker, with prompt caching and compaction built in. Outcomes let a separate grader check work against your rubric and send the agent back for another pass. Multiagent orchestration lets a lead agent delegate to specialists on a shared filesystem.

AgentCore. Harness lets you declare the model, prompt, tools, and memory in configuration, and AWS says you can switch model providers mid-session without losing context. If you want your own loop, Runtime hosts agents built with LangGraph, CrewAI, LlamaIndex, Google ADK, OpenAI Agents SDK, Strands, and others. Evaluations and Optimization add measurement and A/B testing.

Models, cloud, and sandboxes

Claude Managed Agents. Claude models only. Available on the Claude API and on Claude Platform on AWS, which Anthropic operates with AWS handling identity and billing; it is not offered on Bedrock. Self-hosted sandboxes move tool execution into your infrastructure, with guides for AWS Lambda MicroVMs, Cloudflare, Daytona, E2B, Modal, Vercel, and GKE, while orchestration stays with Anthropic.

AgentCore. Model-agnostic: Harness works with Bedrock models, OpenAI, Google Gemini, and OpenAI-compatible providers, and Runtime with any model including Claude, Nova, Llama, and Mistral. Each Harness session runs in its own microVM, and the managed services run in AWS regions. Registry can catalog agents and tools hosted on other clouds or on-premises.

Security, permissions, and memory

Claude Managed Agents. Per-tool permission policies: always allow, always ask (pause for a human), or auto (the server decides). MCP tools default to always ask. Vaults keep tokens out of the session. Every memory change creates an immutable version you can audit, restore, or redact. Anthropic's docs say it is not currently eligible for Zero Data Retention or HIPAA BAA coverage.

AgentCore. Policy intercepts every tool call at the Gateway and checks it against rules written in natural language or a Cedar-compatible policy language. Identity works with Cognito, Okta, Entra ID, and Auth0. Observability uses OpenTelemetry and CloudWatch, inside your existing AWS controls.

Pricing

Claude Managed Agents. Tokens at standard Claude API rates plus $0.08 per session-hour while running, with idle time free. Web search is $10 per 1,000 searches. On Claude Platform on AWS, charges convert to Claude Consumption Units on your AWS bill.

AgentCore. Consumption-based with no upfront commitment. Runtime microVMs are $0.0895 per vCPU-hour and $0.00945 per GB-hour on v1, or $0.1276 and $0.0169 on v2, billed per second. Gateway is $0.005 per 1,000 invocations and Policy is $0.000025 per authorization request. Model usage is billed separately.

Where Runtime fits

Both are excellent foundations for engineering teams, and both leave the same work for a payment company: approvals tied to money movement, PCI and PII stripping, an audit trail an examiner accepts, an interface ops teams use, and independence from a single vendor. Claude Managed Agents ties you to one model vendor; AgentCore ties you to one cloud.

Runtime is the AI agent harness for payment and fintech teams, and it removes both dependencies. It runs in your AWS, GCP, or Azure account or fully self-hosted, with your own sandbox provider and fallbacks, and routes across Anthropic, OpenAI, Google, and open-weight models and across Claude Code, Codex, and OpenCode. If a provider has an outage, changes pricing, or deprecates a model, agents keep running. Guardrails are yours: read-only by default, approvals before money moves, RBAC, masked credentials, card numbers and SSNs stripped from prompts and logs, and every run recorded end to end. It is built for teams, with shared agents, templates, and memory across support, payment ops, finance, risk, and compliance. A forward-deployed AI engineer who has built payment systems maps your processes and builds the first agents with you. An engineering team can spend two quarters building that layer on either platform; Runtime gives it on day one. Rain replaced $250k in vendor spend with Runtime.

Which should you choose

• Choose Claude Managed Agents if you are standardizing on Claude and want the model maker's harness, with outcomes, multiagent orchestration, and versioned memory.
• Choose AgentCore if you are an AWS shop that wants modular, model-agnostic services, a Cedar-compatible policy engine, and everything inside your AWS controls.
• Choose Runtime if payment ops, risk, compliance, or finance need agents on real queues with approvals before money moves, across clouds, sandboxes, and model providers.