Amazon Bedrock AgentCore Review 2026: Pricing, Features and Verdict

What it is

AgentCore is AWS's managed platform for deploying and operating AI agents. It is not an agent builder in the drag and drop sense: you write the agent in LangGraph, CrewAI, Strands, the OpenAI Agents SDK, the Claude Agent SDK or another framework, then use AgentCore services to host it, give it tools, memory and identity, and watch it in production.

The components:

• Runtime: serverless hosting with a dedicated microVM per user session, scaling from zero to hundreds of thousands of concurrent sessions, with bidirectional streaming for real time and voice agents. An instance option runs agents on AWS managed EC2 for up to 14 days.
• Gateway: turns APIs, Lambda functions and existing MCP servers into tools an agent can call, with semantic tool search.
• Memory: short term session memory and long term memory records.
• Identity: agent authentication with Amazon Cognito, Microsoft Entra ID and Okta, and credential management for services like Salesforce, GitHub and Stripe.
• Browser and Code Interpreter: managed cloud browser and sandboxed code execution.
• Web Search, Policy, Evaluations, Observability, Optimization, Agent Registry and Payments.

AWS lists 15 regions, VPC support across the main services and PrivateLink on Runtime. The Amazon Bedrock SLA applies.

Pricing explained

All prices are from AWS's AgentCore pricing page in October 2026. There is no subscription, minimum fee or upfront commitment.

Component	Price
Runtime microVMs v1	$0.0895 per vCPU hour, $0.00945 per GB hour
Runtime microVMs v2	$0.1276 per vCPU hour, $0.0169 per GB hour
Runtime v2 committed baseline (launching by October 2026)	$0.0997 per vCPU hour, $0.0132 per GB hour
Runtime instances	EC2 On Demand cost plus 12% management fee (7.8% on G series)
Browser, Code Interpreter	$0.0895 per vCPU hour, $0.00945 per GB hour each
Web Search	$7.00 per 1,000 queries
Gateway	$0.005 per 1,000 API calls, $0.025 per 1,000 searches, $0.02 per 100 tools indexed a month, $0.006 per GB to your VPC
Identity	$0.010 per 1,000 token or key requests (free via Runtime or Gateway)
Memory, short term	$1.00 per GB ingested, $0.20 per GB retrieved, $0.10 per GB month stored (from October 6, 2026)
Memory, long term	$0.75 per 1,000 records a month ($0.25 self managed), $0.50 per 1,000 retrievals
Policy	$0.000025 per authorization request
Evaluations	$0.0024 per 1,000 input and $0.012 per 1,000 output tokens; custom $1.50 per 1,000 evaluations
Observability	Amazon CloudWatch rates

How Runtime billing works. Per second, 1 second minimum, 128 MB minimum memory. CPU is billed only when active, so the seconds an agent spends waiting on a model response are not billed for CPU. Memory is billed on what the session holds; v2 reclaims idle memory after 120 seconds. AWS's pricing page lists both v1 and v2 rates without saying which applies to a new agent, so check your console.

Model tokens are extra. AgentCore fees cover infrastructure. Claude, Nova, OpenAI or any other model is billed by Bedrock or the model provider on top.

Memory pricing changed. Short term ingestion was $0.25 per 1,000 events through October 5, 2026 and moved to per GB on October 6. Events bill at a 12 KB minimum.

Worked examples

• One active agent hour: 1 vCPU and 2 GB on v2 microVMs for an hour of active compute is $0.1276 + (2 x $0.0169) = $0.1614, before model tokens.
• AWS's support agent example: 1 million Runtime v2 sessions totals $6,703 on AWS's page.
• Research agent: 10,000 web searches is $70 in Web Search fees alone, before compute and tokens.
• HR assistant on Gateway: AWS's example totals $2,250.04 a month.

AWS's own examples range from $1.50 (Identity) to $6,703 (Runtime at 1 million sessions). Budget per component, then add model costs, which are often the largest line. Compare against seat and credit priced platforms in the AI Price Index.

Key features

Session isolation

Every user session gets its own hardware isolated microVM, so one user's agent cannot read another's memory or files. This matters for multi tenant SaaS agents and regulated data.

Long running sessions

Up to 8 hours on microVMs and up to 14 days on instances, enough for coding agents, data pipelines and research tasks that most serverless platforms would time out.

Framework and model freedom

Bring any framework and any model. You are not tied to Bedrock models, which lowers lock in compared with platform specific builders.

MCP and A2A

MCP is supported across AgentCore, and Gateway can front existing MCP servers. Agent to agent (A2A) protocol works on Runtime, with broader support listed as coming soon.

Managed tools

Browser for web automation, Code Interpreter for sandboxed Python and analysis, Web Search, and Gateway for your own APIs. Each is a managed service, so you skip running headless browsers or sandboxes yourself.

Identity and policy

Agents act with scoped credentials through Cognito, Entra ID or Okta, and Policy enforces authorization rules per request, with natural language policy authoring.

Evaluations and observability

Built in and custom evaluations score agent quality, and CloudWatch based observability traces sessions and tool calls.

Setup

AgentCore is code first. A typical path: build the agent locally in your framework, package it, deploy it to Runtime with the AWS CLI or SDK, register tools in Gateway, attach Memory and Identity, and turn on Observability. Expect AWS IAM, VPC and CloudWatch knowledge. There is no visual canvas, so business teams will need engineers.

What users say

We found no G2, Capterra or Trustpilot listing for AgentCore with 10 or more reviews, so there is no rating to report. AWS as a whole has a long track record, and the Bedrock SLA applies, but AgentCore itself is a newer service with a short public track record.

Privacy and security

AgentCore is covered by AWS compliance programs including SOC, PCI, HITRUST, IRAP and the ISO 27001, 27017, 27018, 27701, 22301, 20000 and 9001 standards, and AWS lists it as HIPAA eligible. FedRAMP is in progress. VPC and PrivateLink keep traffic off the public internet. The FAQ does not address training on customer data; for Bedrock hosted models, check the model's data terms, and for third party models, the provider's.

Who it is for

• Engineering teams already on AWS shipping customer facing or internal agents.
• SaaS companies that need per user session isolation.
• Regulated teams that want HIPAA eligible infrastructure in their AWS account.
• Teams running long agent jobs that exceed typical serverless timeouts.

Who should skip it

• Business teams without engineers: look at Lindy, n8n or Microsoft Copilot Studio.
• Teams on Google Cloud or Azure, where native platforms fit better.
• Anyone who needs a fixed, predictable monthly bill.
• US federal workloads that require FedRAMP today.

Alternatives

• OpenAI AgentKit review: agent builder and SDK with no platform fee, OpenAI models.
• Microsoft Copilot Studio review: low code agents for Microsoft 365, credit priced.
• IBM watsonx Orchestrate review: enterprise agent orchestration from $530 a month.
• Salesforce Agentforce review: agents inside Salesforce at $0.10 per action.
• Gemini Enterprise review: Google's agent platform for Workspace and Google Cloud.
• More in best AI agent builders and best enterprise AI agent platforms.

Verdict

If your engineers build on AWS, AgentCore is the most complete way to run agents in production: isolated sessions, long runtimes, any model, MCP and AWS compliance, with CPU billed only when it works. Price each component and the model tokens before you launch, set CloudWatch budgets, and keep a test harness for cost per session. If you do not have engineers, it is the wrong tool.