Amazon Bedrock AgentCore: modular agent infrastructure on AWS
What does Amazon Bedrock AgentCore provide?
Amazon Bedrock AgentCore is a modular platform for running and connecting agents, securing their access to tools, and monitoring and improving their behavior. AWS describes framework and model flexibility rather than requiring one proprietary agent implementation. Individual capabilities can be used independently or together. Runtime, Gateway, Memory, Identity, Policy and observability have separate roles. Published consumption rates make unit-cost estimation possible, but there is no single universal cost per agent or conversation. Source Source
How do Runtime deployment and billing options differ?
The current pricing documentation describes direct code deployment and container-based deployment, and two compute types: serverless microVMs and managed EC2 instances. microVM billing is calculated per second with a one-second minimum. CPU consumption can fall to zero during I/O waits, while memory treatment depends on the runtime version; v2 automatically reclaims idle memory after 120 seconds. Charges include system overhead, and a 128 MB minimum memory bill applies. Source
Instance-based Runtime is for persistent, resource-intensive, or specialized workloads. AWS bills the underlying EC2 instance plus an AgentCore management fee calculated against public EC2 On-Demand pricing. Existing EC2 discounts can apply to the compute charge, not the management fee. The page describes sessions up to 14 days and multiple agents sharing one instance. EBS storage remains separately billable, including while a session is stopped. Container artifacts use separately billed ECR storage; direct-code artifacts are billed at S3 Standard rates. Standard network-transfer charges are another separate cost dimension. Source
How do Gateway, protocols and Identity work?
Gateway transforms APIs and Lambda functions into agent-compatible tools and connects existing MCP servers. Runtime's A2A documentation specifies an HTTP server on port 9000 and an Agent Card at /.well-known/agent-card.json. AWS supports IAM SigV4 or OAuth 2.0 authorization for that deployment contract. Teams still configure their agents' logic, contracts, authorization, and task delegation. Source Source
Identity manages access to AWS resources and third-party services on behalf of agents or users, including preauthorized user consent. The pricing distinction is especially useful: Identity adds no separate Identity charge when used through Runtime or Gateway, while other scenarios are metered by successful requests for OAuth tokens or API keys for non-AWS resources. Policy has a separate authorization-request meter and an input-token meter. Source
How are memory, observability and evaluations billed?
Memory separates short-term event data from long-term extracted memories. The current price page charges short-term ingestion, retrieval, and storage separately, with minimum and maximum billable sizes per ingested or retrieved event. Long-term storage uses record counts, and retrieval uses request counts. Built-in extraction and override/self-managed extraction have different storage rates, with possible model charges in the customer's account for the latter. Retention and event size therefore matter to cost; counting agent sessions alone is insufficient to estimate the memory bill. Source
Observability stores telemetry in CloudWatch, where ingestion, retention, queries, and sensitive-data masking can generate charges. Evaluations offers built-in and custom evaluators; model usage is included in built-in evaluator pricing, while custom evaluations have separate model-inference costs. The live page also lists Optimization: Insights is free during public preview, recommendation generation has no charge, and experiments consume underlying AgentCore services. These terms do not make the surrounding telemetry, evaluation, or runtime consumption free. Preview terms can change. Source
How should you estimate an AgentCore bill?
Build the estimate from session CPU and memory, selected tools, gateway invocations and indexing, identity requests outside included paths, memory ingestion and retention, retrieval calls, policy checks, evaluation sampling, and telemetry. Add model inference and surrounding AWS infrastructure when separately billed. Registry free tiers also apply to distinct meters, so staying below 5,000 records does not guarantee a free registry if Search or List/Get traffic exceeds its separate allowance. Source
The price page describes pay-as-you-go usage without upfront commitments or minimum fees, while separately listing an optional future-labeled Runtime commitment option. It also advertises up to $200 in Free Tier credits for new AWS customers, not an unconditional perpetual allowance for every customer. An estimate should emphasize modular metering and implementation-specific totals rather than naming one headline monthly price.
What should an AgentCore pricing pilot record?
Measure the units AWS bills, not only the number of chat sessions. Capture active CPU and memory consumption, gateway requests, memory event sizes and retention, identity requests outside included paths, policy checks and telemetry. Separate model inference, storage and network charges where the pricing page does. This is an estimation method, not a benchmark or an invented typical monthly bill.
Compare the resulting estimate with other infrastructure layers, the protocols guide and the RFP checklist. BAND (band.ai) lists Amazon Bedrock AgentCore and Strands Agents as integrations and supplies shared rooms for cross-vendor work with people. Keep AWS runtime cost and any collaboration-service quote separate. Confirm the actual integration and Enterprise requirements before presenting the combined architecture as production-ready.
| Component | Published rate | Qualification |
|---|---|---|
| Runtime v1 | $0.0895/vCPU-hour + $0.00945/GB-hour | Consumption and surrounding charges |
| Runtime v2 | $0.1276/vCPU-hour + $0.0169/GB-hour | Consumption pricing |
| Runtime instances | EC2 cost plus management fee | Check instance-specific terms |
| Gateway | $0.005/1,000 invocations | Search $0.025/1,000; indexing separate |
| Identity | $0.010/1,000 successful requests | No extra Identity charge through Runtime/Gateway |
| Short-term Memory | $1.00/GB ingestion; $0.20/GB retrieval; $0.10/GB-month storage | Effective 6 October 2026; event-size bounds apply |
| Long-term Memory | $0.75/1,000 records/month built-in | Override/self-managed $0.25; retrieval $0.50/1,000 |
| Policy | $0.000025/authorization request | Input tokens $0.13/1,000; other safeguards separate |
| Agent Registry | 5,000 records/month free, then $0.40/1,000 | Search and List/Get have separate meters |
| Observability | CloudWatch rates | Ingestion, retention and queries apply |
What does Amazon Bedrock AgentCore do well?
- Framework- and model-agnostic.
- Broad protocol support on Runtime: HTTP, MCP, A2A, AG-UI.
- The most detailed public pricing in this guide.
What are the limitations of Amazon Bedrock AgentCore?
- Agent infrastructure; application experience and human workflow depend on the implementation.
- Runs on AWS; no non-AWS runtime is stated.
- Many metered components make cost forecasting work.
How does Amazon Bedrock AgentCore fit with BAND (band.ai)?
AgentCore hosts the agents; BAND is where they work with agents outside AWS and with people. BAND lists Amazon Bedrock AgentCore and Strands Agents as integrations, so AgentCore agents can join BAND rooms without being rebuilt.
Where do the facts about Amazon Bedrock AgentCore come from?
- aws.amazon.com/bedrock/agentcore
- aws.amazon.com/bedrock/agentcore/pricing
- docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-a2a.html
Updated 3 November 2026
What do buyers ask about Amazon Bedrock AgentCore?
Is there one fixed AgentCore monthly price?
No. AWS publishes separate consumption meters for the capabilities used. Add model and surrounding infrastructure charges where applicable.
What are Runtime v1 rates?
The pricing page lists $0.0895 per vCPU-hour and $0.00945 per GB-hour. These are compute meters, not a complete application cost.
What changed in short-term Memory pricing?
The page lists GB-based ingestion, retrieval and storage rates effective 6 October 2026. The old $0.25 per 1,000 new events applied through 5 October.
Is AgentCore Identity always free?
There is no additional Identity charge through Runtime or Gateway. Other successful requests for non-AWS OAuth tokens or API keys are metered.
Does a registry free tier cover all registry usage?
No. Record counts, Search calls and combined List/Get calls have distinct free allowances and paid meters.
How does Amazon Bedrock AgentCore compare with other vendors?
Connect AgentCore agents with agents from other vendors