---
title: "AWS Bedrock AgentCore alternatives in 2026: Claude Managed Agents, LangSmith Deployment, OpenAI Agents API, and Cadenya"
url: "https://cadenya.com/other-content/aws-bedrock-agentcore-alternatives"
description: "AWS Bedrock AgentCore alternatives in 2026: Runtime session duration (8 hours), idle timeout (15 minutes by default), approval hooks (900 seconds at most), pricing per vCPU-hour, Gateway OpenAPI targets, Code Interpreter, and regions, compared with Claude Managed Agents, LangSmith Deployment, OpenAI's Agents API, and Cadenya."
---

Amazon Bedrock AgentCore is AWS’s kit for running the agent you write: each session gets its own microVM, and you add Memory, Gateway, Identity, a Code Interpreter, and a Browser as you need them. It runs any framework and any model, and it carries the longest compliance list in this category.

Its limits are about time. A session lasts “up to 8 hours on microVMs.” One that sits idle ends after 15 minutes by default. And the built-in approval hook waits at most 900 seconds before it denies. So the usual reason to look elsewhere is an agent that has to wait, or a team that doesn’t want to write the loop at all.

Every fact below comes from the vendors’ own docs and pricing pages, checked on September 28, 2026.

## What AgentCore does well

-   **Any framework, any model.** Runtime “works with custom frameworks and any open-source framework, including CrewAI, LangGraph, LlamaIndex, Google ADK, OpenAI Agents SDK, and Strands Agents”, and with models from Bedrock, OpenAI, Gemini, Claude, and more.
-   **Compliance.** “HIPAA eligible and FedRAMP (Class C and Class D), SOC 2 and ISO”, plus VPC and PrivateLink connections.
-   **Pay per second of use.** Runtime is $0.0895 per vCPU-hour and $0.00945 per GB-hour on v1 microVMs, with “no upfront commitments or minimum fees.” AWS’s worked example, a support agent with 1 million sessions a month, comes to $6,703.
-   **Parts you’d otherwise build.** With Gateway, an OpenAPI target (or a Smithy model, or a Lambda function) becomes MCP tools ($0.005 per 1,000 calls). The Code Interpreter runs Python, JavaScript, and TypeScript in a sandbox. Memory stores events and long-term records.
-   **Regions.** Runtime, Gateway, and the built-in tools (Code Interpreter and Browser) run in all seven European regions on AWS’s table, including Frankfurt, Ireland (eu-west-1), London (eu-west-2), and Paris (eu-west-3), and in AWS GovCloud (US-West). The harness and Memory aren’t in Milan or Spain.

## Where it stops

1.  **Sessions end.** The maximum session duration is 8 hours on microVMs (14 days on the EC2-backed Instances). The idle timeout defaults to 900 seconds, “15 minutes”, and goes up to 8 hours.
2.  **Approvals are short.** A harness hook calls a Lambda function for an allow or deny decision, and its timeout “defaults to 60 seconds and accepts values from 1 through 900 seconds”, with deny as the default on timeout.
3.  **You write the loop.** AgentCore hosts your agent’s code. The loop, the prompts, the tool routing, and the stream to your UI stay yours. (The Strands-based harness is the exception, and it’s the one with the approval hook.)
4.  **Sessions and users are yours to map.** AgentCore doesn’t track which session belongs to which user. AWS says “your client backend should maintain the relationship between users and their session IDs.”
5.  **AWS only.** No page describes running it outside AWS.

## The alternatives

So where do you go when the wait is longer than a session, or you don’t want to write the loop?

### Claude Managed Agents

Anthropic runs the loop and the sandbox, and approvals set to `always_ask` wait “indefinitely.” You pay $0.08 per session-hour only while the session is running, plus tokens. The trade: Claude models only, one model per session, a beta, and “not currently eligible for Zero Data Retention or HIPAA”, which is the opposite of AgentCore’s strongest point. [Claude Managed Agents alternatives](/other-content/claude-managed-agents-alternatives) covers it in depth.

### LangSmith Deployment

Also hosts the agent you write, but its approvals are LangGraph interrupts, which wait “indefinitely until you resume execution.” Deployments run in LangChain’s cloud in the US or the EU. Plus is $39 a seat a month plus metered runtime at $0.0675 per vCPU-hour. It publishes no compliance statement in its docs, and self-hosting needs the Enterprise plan.

### OpenAI’s Agents API

A managed harness from OpenAI, in beta, with sessions, compaction, and recovery, and the model can change between turns. It’s US-only for data residency with no Zero Data Retention, and it has no approval step of its own. Pricing is model tokens, tools, and $0.03 per 20 minutes for a 1 GB sandbox.

### Cadenya

Cadenya runs the loop for you, so there’s no agent code to deploy. You point a tool set at your OpenAPI document, pick any model from a provider you connect, and start an objective, with approvals set per tool set. An objective can wait up to 24 hours for an approval, against AgentCore’s 15-minute idle default. It’s free for 1,000 loops a month and $49 for 50,000. [What it offers that AgentCore doesn’t](#what-does-cadenya-offer-that-agentcore-doesnt) is at the end.

### Your own loop on a durable engine

If the waits are long and the agent must run in your own cloud, run the loop yourself on Temporal, Restate, DBOS, or Inngest. All four can hold an agent through days of waiting. [DBOS vs Temporal](/other-content/dbos-vs-temporal-for-ai-agents) is a good first read if your app already runs on Postgres.

## Side by side

|  | Amazon Bedrock AgentCore | Claude Managed Agents | LangSmith Deployment | OpenAI Agents API | Cadenya |
| --- | --- | --- | --- | --- | --- |
| Who writes the loop | You | Anthropic | You | OpenAI | Cadenya, configured by you |
| Longest session | 8 hours on microVMs | Not stated | Not stated | Not stated | Not stated |
| Idle | 15 minutes by default, up to 8 hours | Not billed | Not stated | Sandbox deleted after an hour | Up to 24 hours, set per variation |
| Approval wait | 1 to 900 seconds | Indefinite | Indefinite | No approval step | Until answered, within the idle timeout you set |
| Models | Any | Claude | Any your code calls | OpenAI | Any provider you connect |
| Compliance lines | HIPAA, FedRAMP, SOC 2, ISO | No ZDR or HIPAA | None found | No ZDR | Not stated |
| Starting price | Per second, no minimum | $0.08 per running hour, plus tokens | $39 a seat a month, plus runtime | Tokens and sandbox time | $0 for 1,000 loops |

## When to stay on AgentCore

Stay if you’re on AWS, you write the agent anyway, the compliance list matters, and a session’s work, including any approval, fits in its limits. For many chat and tool agents, it does.

## What does Cadenya offer that AgentCore doesn’t?

AgentCore hosts the agent you write, on AWS. **Cadenya runs the agent for you: the loop, its durable state, and its human-in-the-loop approvals, with no code to deploy.** Here’s what that covers:

1.  **The agent loop, already written.** AgentCore Runtime runs your code in a microVM per session, and the loop, the prompts, the tool routing, and the stream to your UI stay yours. On Cadenya you configure the agent on a variation, and Cadenya calls the model, runs the tools, streams each reply, and compacts the context window when a conversation gets long.
2.  **One service instead of a kit.** There’s no agent code to deploy, no vCPUs or memory to size, and no AWS account to set up. Where AgentCore splits the job into Runtime, Memory, Gateway, and Identity, Cadenya is one API.
3.  **Approvals that wait a day, not 15 minutes.** An objective can sit idle waiting for an approval for up to 24 hours. AgentCore’s idle timeout defaults to 15 minutes, and its approval hook waits 900 seconds at most before it denies.
4.  **Human-in-the-loop approvals, screen included.** A tool marked for approval pauses the objective on `toolApprovalRequested` until your backend, a webhook handler that posts to Slack, or a person in the React chat widget approves or denies it. A denial can carry a memo that steers the agent.
5.  **Users mapped to conversations.** AgentCore “does not enforce session-to-user mappings”, so your backend tracks them. On Cadenya, your backend mints a widget session per user, and Cadenya derives the tenant, the user, the secrets, and the pinned parameters from it. The session can carry the user’s short-lived token as a secret that overrides the shared API credential, so your API sees the signed-in user on each tool call.
6.  **Your OpenAPI document as the tools, with no per-call fee.** Gateway turns an OpenAPI target into MCP tools at $0.005 per 1,000 calls. A Cadenya tool set turns every operation into a tool, re-checks the document each hour, filters the list, and loads tools on demand, and tool calls aren’t billed.
7.  **A price in one line.** AgentCore bills vCPU-hours and GB-hours by the second, plus Gateway, Memory, and the built-in tools as you use them. Cadenya bills per loop, one LLM request: free for 1,000 a month, $49 a month for 50,000, then $0.0015 each.

[How objectives work](/docs/guides/the-basics/objectives) walks through an agent turn on Cadenya, and the free plan needs no card.