AI models and agent platforms · Agent build and run platforms
Amazon Bedrock AgentCore
Modular AWS services to run, connect, secure and observe production agents, with any framework or model.
- Vendor
- AWS
- Pricing model
- Consumption-based per component (runtime, gateway, memory, browser, code, policy), plus model tokens.
- Runs on
- Your AWS account and regions
- Agentic OS layer
- Agents
- Industries
- Any industry
- Last reviewed
- November 2026
Our verdict
The strongest choice when agents must run inside AWS with enterprise identity, policy and observability. It is infrastructure, not a product for business users: you need engineers, and you need to watch a bill with many small meters.
What it is
AgentCore is AWS's set of building blocks for running agents in production. You bring an agent written in any framework, such as LangGraph, CrewAI or AWS's own Strands, and any model, and AgentCore supplies the parts that are hard to build safely: an isolated runtime, a gateway that turns APIs and Lambda functions into agent tools, identity, memory, a browser, a code interpreter, policy controls, evaluations and observability.
Each component can be used on its own and is billed on use, with no upfront commitment. AWS put the older Bedrock Agents product into maintenance for new customers after July 2026, so new agent builds on AWS should start on AgentCore.
Best for
Organisations with AWS as their main cloud and engineering teams that need agents to call internal systems securely, with audit trails and policy controls.
Not the right fit
Business teams that want to build agents without code, and organisations with no AWS footprint.
Key capabilities
Runtime
Isolated micro-VM sessions for each agent run; CPU is billed only while active, not while waiting on the model.
Gateway
Turns APIs, Lambda functions and existing MCP servers into tools agents can call securely.
Identity
Agent and user identity for calling third-party and internal systems, included with runtime or gateway.
Memory
Short- and long-term memory so agents keep context across sessions.
Browser and code interpreter
Managed tools for web tasks and safe code execution.
Policy, evaluations and observability
Boundaries on what agents may do, quality evaluations, and traces in CloudWatch.
Strengths
- Framework- and model-agnostic
- Security, identity and policy built for regulated workloads
- Pay only for what runs; idle waiting is cheap
- Native fit with AWS data, Lambda and networking
Watch-outs
- Twelve-plus separate meters make cost forecasting work
- Needs engineers; there is no business-user builder
- Model token cost usually dwarfs the AgentCore lines
- Fixed network costs, such as NAT gateways, bill even when nothing runs
Pricing
AgentCore has no upfront fees or minimums. Runtime, browser and code interpreter bill on vCPU-hours and GB-hours of memory during a session. Gateway bills per tool call and search; memory bills on data ingested and stored; observability bills at CloudWatch rates. Identity is free when used through runtime or gateway. New AWS customers can get free-tier credits.
Model calls through Bedrock bill separately per token and are usually the largest line. Rates vary by region and change often, so model costs with AWS's pricing calculator before you commit.
Checked November 2026. Vendors change plans often, so confirm current pricing with AWS before you budget.
Security and governance
Runs in your AWS account with IAM, VPC connectivity, session isolation in micro-VMs, policy controls on agent actions, and logs and traces in CloudWatch. Data stays in the regions you choose.
How we would roll it out
- Choose the frameworkPick the agent framework your engineers know; AgentCore runs most of them.
- Wrap tools in the gatewayExpose the APIs agents need through the gateway with scoped identity, rather than giving agents raw credentials.
- Set policy and evaluationsDefine what each agent may do and evaluate output quality before go-live.
- Tag and budgetTag every agent's resources and set AWS budgets per agent, including model spend.
Compare with
Other tools in AI models and agent platforms, closest match first.
Microsoft Copilot Studio
Agent build and run platforms
Low-code agent builder for Microsoft 365, Teams, Dynamics and Power Platform, billed on Copilot Credits.
Best for Microsoft 365 organisations building internal and customer-facing agents.
LangGraph and LangSmith
Agent build and run platforms
Open-source framework for controllable, stateful agents, with LangSmith for tracing, evaluation and deployment.
Best for Engineering teams that want full control of agent logic and to stay cloud-neutral.
ChatGPT Enterprise
Enterprise AI assistants
The most widely adopted general AI assistant, with enterprise admin, residency and compliance controls.
Best for Organisations that want one assistant everyone already knows how to use.
Claude Enterprise
Enterprise AI assistants
Assistant, coding agent and desktop task agent in one seat, with published seat pricing and usage billed separately.
Best for Teams that work with long documents, write a lot, or want coding and desktop agents in the same seat.
Gemini Enterprise
Enterprise AI assistants
Google's agent platform for every employee: enterprise search, prebuilt agents and a no-code agent builder.
Best for Google Workspace organisations that want search and agents across company data.
Choosing between Amazon Bedrock AgentCore and the alternatives?
Book a call. We look at your systems, data and governance together and tell you which platform fits, and what to build on it first.