Amazon Bedrock Agents is the agent orchestration service within Amazon Web Services' broader Bedrock platform, which reached general availability in September 2023. Where open-source frameworks like CrewAI and LangGraph require teams to build and host their own infrastructure, Bedrock Agents provides a fully managed runtime that handles orchestration, memory, guardrails, and code execution out of the box.

The commercial traction is substantial. Bedrock has grown into a multi-billion dollar annualized run rate business, with customer spend up 60 percent quarter over quarter in Q4 2025 and a customer base that grew 4.7x year over year. Named customers span industries: Pfizer estimates AI workloads on Bedrock will save $750 million to $1 billion annually, NatWest Group uses the platform to combat financial crime, and BMW is building its AI assistant on the service.

In October 2025, AWS launched Amazon Bedrock AgentCore, adding modular services for runtime, memory, identity, policy controls, observability, and evaluation. A partnership with OpenAI to build a stateful runtime environment for persistent agents signals that AWS sees agent infrastructure as a core battleground.

Enterprise compliance that is hard to match. Amazon Bedrock inherits the deepest compliance portfolio in cloud computing. The service is in scope for SOC 1, 2, and 3 (Type II), ISO 27001, ISO 27017, ISO 27018, CSA STAR Level 2, is HIPAA eligible, GDPR compliant, and FedRAMP High authorized in GovCloud. For regulated industries, from healthcare to financial services to government, no other AI agent platform offers this breadth of certifications at launch.

Guardrails are a first-class product, not an afterthought. Amazon Bedrock Guardrails provides six configurable safeguard policies: content filters, denied topic classification, PII redaction, word filters, prompt attack detection, and hallucination detection via automated reasoning checks. These apply across individual inference calls, agent workflows, knowledge base queries, and multi-node flows. As we explored in our analysis of the agentic commerce stack, governance is the missing layer for most agent deployments. Bedrock ships it natively.

Model-agnostic by design, with depth across the catalogue. Bedrock Agents supports Claude, GPT-4.1, Llama, Mistral, Cohere, and Amazon's own Nova models through a unified API. Teams can switch foundation models without rewriting agent logic. The platform now offers over 100 foundation models, making it one of the broadest AI model marketplaces available.

Multi-step orchestration with knowledge bases and code execution. Agents break down user requests into logical sequences, invoke action groups (APIs), query knowledge bases for retrieval-augmented generation, and execute code in a secure sandbox. AgentCore adds long-term memory, policy controls, and a suite of 13 pre-built evaluation systems for monitoring correctness, safety, and tool selection accuracy.

Complexity and cost opacity beneath the managed surface. Bedrock Agents uses pay-per-use pricing based on underlying foundation model tokens, knowledge base queries, and guardrail evaluations. A single user interaction can trigger multiple internal model calls, making actual costs significantly higher than naive estimates. There is no published per-agent price, no free tier for agents, and cost prediction requires careful instrumentation. For teams used to simple SaaS pricing, this is a learning curve.

AWS lock-in is the trade-off for managed convenience. Bedrock Agents runs on AWS infrastructure and integrates tightly with IAM, Lambda, S3, and VPC. Migrating agent workflows to another cloud or to self-hosted infrastructure would require significant re-engineering. Organisations that prioritise multi-cloud portability should weigh this against the operational simplicity.

Amazon's own agent challenges raise questions. In December 2025, Amazon's internal AI coding assistant Kiro deleted a production environment and caused a 13-hour AWS outage. Four major AI-related incidents in three months led Amazon to mandate senior approval for AI-assisted code changes. The Bedrock Agents product itself was not implicated, but the episodes highlight that even the platform provider is still learning how to govern autonomous agents in production.

Amazon Bedrock Agents operates on consumption-based pricing with no upfront commitments. You pay for foundation model inference (per input and output token), knowledge base queries, guardrail evaluations, and AgentCore services (billed per second of CPU and memory consumption). AgentCore pricing covers runtime, memory, gateway, policy, identity, observability, and evaluation modules independently. Deployment is cloud-native within AWS regions, with VPC connectivity and PrivateLink support. No self-hosted option exists.

Amazon Bedrock holds SOC 1, 2, and 3 (Type II), ISO 9001, ISO 27001, ISO 27017, ISO 27018, ISO 27701, CSA STAR Level 2, and is HIPAA eligible, GDPR compliant, and FedRAMP High authorized in AWS GovCloud. Complete session isolation, VPC connectivity, and PrivateLink support provide enterprise-grade network security. Data is not used to train models.

Overall: 3.5/5

Amazon Bedrock Agents is the right choice for enterprises already invested in AWS that need production-grade AI agents with the deepest compliance coverage available. If your organisation operates in healthcare, financial services, or government, and regulatory certification is non-negotiable, Bedrock Agents delivers what open-source frameworks cannot match out of the box. Teams that want multi-cloud portability, transparent pricing, or rapid prototyping without AWS expertise should evaluate CrewAI or LangGraph first. With a multi-billion dollar run rate, a 360-partner ecosystem, and the OpenAI stateful runtime partnership, AWS is building the enterprise default for governed AI agents. The question is whether managed convenience justifies the cost opacity and platform lock-in.

Try Amazon Bedrock Agents: aws.amazon.com/bedrock/agents

Can the company that struggled to govern its own internal AI agents build the governance layer the rest of the enterprise needs?

Editorial disclaimer: Reviews reflect the independent editorial assessment of Major Matters and are not sponsored or endorsed by the companies reviewed. We recommend conducting your own evaluation to determine whether any product is the right fit for your specific requirements.