Amazon Bedrock AgentCore
Source: AWS | Type: Vendor | Category: ai-ml / ai-agent-platform
What It Does
Amazon Bedrock AgentCore is AWS’s managed platform for building and operating AI agents in production, announced at re:Invent 2025 and progressively reaching general availability through Q1 2026. It bundles the infrastructure concerns of agent deployment — code execution sandboxes, browser automation, session memory, identity federation, observability traces, and policy enforcement — into a single AWS-integrated service.
AgentCore is designed as a complement to whatever agent framework a team already uses (LangGraph, LlamaIndex, Agno, etc.) rather than as a competing orchestration layer. It handles the “outer shell” of agent deployment: where the agent runs, how it talks to tools, how its actions are constrained and logged, and how its memory persists across sessions. Individual components (Runtime, Gateway, Memory, Identity, Browser, Code Interpreter, Observability, Evaluations, Policy) can be adopted incrementally.
Key Features
- Runtime: Serverless, secure execution environment for agent logic; no infrastructure management required; scales to zero and back
- Gateway: Unified tool access layer with MCP server support and server-side tool execution for 100+ preconfigured integrations; connects agents to AWS services and third-party APIs
- Memory: Persistent cross-session context with episodic and semantic memory; agents learn from past interactions without requiring client-side state management
- Identity: Seamless authentication for agents across AWS services and third-party APIs; handles credential delegation without exposing keys to agent code
- Browser: Sandboxed browser with OS-level automation — mouse clicks, keyboard input, drag-scroll, screenshot capture at OS coordinates (April 2026 GA)
- Code Interpreter: Sandboxed code execution environment for agent-generated code; multi-language support
- Observability: Distributed tracing, logging, and debugging for agent runs across all AgentCore services
- Evaluations: Continuous quality scoring with customizable metrics; GA March 2026; integrates human feedback loops
- Policy: Fine-grained controls over what tools agents can invoke and with what parameters; GA March 2026
Use Cases
- Enterprise agent deployment on AWS: Organizations already on AWS wanting managed infrastructure for production agents without building sandbox, memory, and gateway infrastructure themselves
- Regulated-industry agent automation: Policy and Identity components address compliance requirements (HIPAA, SOC 2) for agents operating on customer data
- Browser automation at scale: AgentCore Browser enables agent-driven web task automation (form completion, data extraction, UI testing) without managing browser fleets
Adoption Level Analysis
Small teams (<20 engineers): Technically accessible but likely cost-prohibitive relative to open-source alternatives (E2B, Daytona) for non-AWS-committed teams. The AWS cost model (per-invocation, per-memory-operation, per-evaluation) can surprise small teams at scale.
Medium orgs (20–200 engineers): Strong fit for AWS-committed teams. The managed infrastructure eliminates the need for a dedicated platform team to build agent sandboxing and observability. Pre-existing AWS relationships simplify procurement and compliance.
Enterprise (200+ engineers): Primary target. Deep AWS integrations (IAM, VPC, CloudTrail, CloudWatch), enterprise support tiers, and the Policy component for governance make this the most natural choice for enterprise organizations already on AWS.
Alternatives
| Alternative | Key Difference | Prefer when… |
|---|---|---|
| E2B | Open-source Firecracker microVMs, faster cold starts, not AWS-specific | Cloud-agnostic or AWS-independent sandboxing with lower per-execution cost |
| Modal | Serverless GPU-native Python, simpler pricing model | ML workloads requiring GPU access without full agent platform overhead |
| LangGraph + custom infra | More control over agent orchestration, open-source | Need custom orchestration patterns not supported by AgentCore’s opinionated model |
| Northflank | Enterprise VPC deployment with GPU, not AWS-native | Multi-cloud VPC isolation with GPU requirements |
Evidence & Sources
- AWS AgentCore official documentation
- Introducing Amazon Bedrock AgentCore, AWS Blog
- AgentCore Evaluations GA, AWS what’s new
- AgentCore Browser OS-level interactions, AWS what’s new
- AWS announces new capabilities for its AI agent builder, TechCrunch
Notes & Caveats
- AWS lock-in: AgentCore is deeply integrated with IAM, VPC, CloudWatch, and other AWS services. Migrating agent infrastructure off AgentCore to a different provider would require rebuilding all gateway integrations, memory backends, and observability pipelines.
- Pricing complexity: The multi-component architecture means cost estimation requires accounting for runtime invocations, memory operations, gateway calls, browser session time, code interpreter compute, and evaluation runs separately. No simple per-agent pricing model.
- Framework agnostic in theory: AWS claims AgentCore works with any framework (LangGraph, Agno, etc.), but practical integration complexity varies. Best-supported path is likely through AWS SDK wrappers and Bedrock native models.
- GA progression: Not all components reached GA simultaneously. Some features were in preview for months before GA. Production teams should verify the GA status of specific components before committing.
- Bedrock model dependency: While AgentCore is framework-agnostic, it naturally integrates with Amazon Bedrock-hosted models. Using external model providers (Anthropic direct, OpenAI) adds integration complexity relative to using Bedrock-hosted model variants.
- No open-source core: Unlike E2B or Daytona, there is no open-source core to evaluate, fork, or self-host. The product is entirely proprietary and managed.