Building AI agents that actually work in production has always been the hard part. You pick a model, write some prompts, and then spend weeks building the infrastructure around it — session management, tool routing, rate limiting, observability, and graceful degradation when things break. On April 22, 2026, AWS shipped an update to Amazon Bedrock AgentCore that eliminates most of that plumbing.

The new managed agent harness and AgentCore CLI let you stand up a working autonomous agent through configuration rather than orchestration code. Three API calls. No backend scaffolding. The harness itself is free — you pay only for the compute your agent actually consumes, billed per-second with zero charges during I/O wait.

This guide covers everything you need to go from zero to a deployed agent: CLI installation, harness architecture, prebuilt coding skills, the MCP Gateway for tool connections, pricing breakdown, and production patterns we recommend at Lushbinary.

1What Changed on April 22, 2026

AWS introduced three major additions to Amazon Bedrock AgentCore:

• Managed Agent Harness — A pre-built infrastructure layer that handles orchestration, compute provisioning, tool connections, and session state. You configure an agent and invoke it; the harness does the rest.
• AgentCore CLI — An npm-distributed command-line tool that scaffolds agent projects, deploys them to AgentCore Runtime, and invokes them directly from your terminal.
• Prebuilt Coding Skills — Ready-to-use agent capabilities for code generation, review, and refactoring that plug into the harness without custom development.

2AgentCore Architecture Overview

AgentCore is structured as a set of composable services you can use independently or together:

The key insight: the harness sits between your code and the runtime, handling all the orchestration logic that teams typically spend weeks building. You define what the agent should do; AgentCore handles how it runs.

3Installing & Using the AgentCore CLI

The CLI is distributed as an npm package. Prerequisites: Node.js (any recent LTS), Python 3.10+, and configured AWS credentials.

The init command generates a Python project with a pre-configured agent definition, tool bindings, and deployment configuration. The generated code uses the AgentCore SDK to define agent behavior:

4The Managed Agent Harness

The harness is the centerpiece of this update. It's a pre-built orchestration layer that handles:

• Session management — Maintains conversation state, tool call history, and context across invocations without you managing a database
• Tool orchestration — Routes tool calls to the right MCP servers, AWS services, or shell commands with automatic retry and error handling
• Guardrails — Enforces token limits, step limits, and timeout boundaries to prevent runaway agents
• Model routing — Supports any Bedrock model or custom inference profile, with the ability to switch models mid-session

The "3 API calls" workflow looks like this:

5AgentCore Runtime: Per-Second Billing

The runtime is where your agent code actually executes. It's a serverless environment built on microVMs with a billing model designed specifically for agentic workloads:

This matters because agentic workloads typically spend 30-70% of their time in I/O wait — waiting for LLM responses, API calls, or database queries. With traditional compute (EC2, Lambda), you pay for that idle time. With AgentCore Runtime, idle time is free.

6MCP Gateway & Tool Connections

The AgentCore Gateway provides a centralized layer for managing how AI agents connect to tools and MCP servers across your organization. It consolidates authentication, observability, and policy enforcement into a single endpoint.

• Centralized auth — OAuth 2.0 Authorization Code flow for MCP server connections, managed at the gateway level
• Policy enforcement — Define which agents can access which tools, with IAM-based access control
• Observability — Every tool call is logged with latency, success/failure, and token usage
• Rate limiting — Protect downstream services from agent-driven traffic spikes

You register MCP servers with the gateway once, and any agent in your account can use them (subject to IAM policies). No per-agent configuration of tool endpoints.

7Prebuilt Coding Skills

The April update also shipped prebuilt coding skills that plug directly into the harness. These give agents the ability to:

• Execute shell commands in a sandboxed environment (launched March 17, 2026)
• Read and write files with persistent session storage (launched March 25, 2026)
• Generate, review, and refactor code using the agent's model
• Run test suites and interpret results

Combined with the managed harness, this means you can deploy a coding agent that clones a repo, makes changes, runs tests, and submits a PR — all without writing orchestration code yourself.

8Pricing Breakdown

AgentCore uses consumption-based pricing with no upfront commitments:

New AWS customers receive up to $200 in Free Tier credits. The "pay only for active compute" model makes AgentCore particularly cost-effective for agents that do a lot of waiting (which is most of them).

9Production Deployment Patterns

Based on our experience deploying AI agents for clients, here are the patterns that work best with AgentCore:

Pattern 1: DevOps Incident Responder

An agent that monitors CloudWatch alarms, investigates incidents by correlating logs and metrics, and posts findings to Slack. Uses shell commands to run diagnostic scripts and the MCP Gateway to connect to PagerDuty and Jira.

Pattern 2: Code Review Agent

Triggered by GitHub webhooks, this agent clones PRs, runs static analysis, executes tests, and posts review comments. The prebuilt coding skills handle file I/O and shell execution; the harness manages the multi-step workflow.

Pattern 3: Customer Support Triage

An agent connected to Zendesk via MCP that reads tickets, queries internal documentation, drafts responses, and escalates complex issues to humans. Session storage maintains context across multi-turn conversations.

10AgentCore vs LangGraph vs CrewAI

How does AgentCore compare to popular open-source agent frameworks?

AgentCore's sweet spot is teams already on AWS who want managed infrastructure without vendor lock-in on the model layer. If you need complex graph-based workflows with conditional branching, LangGraph gives you more control. If you need multi-agent collaboration patterns, CrewAI is purpose-built for that. For a deeper comparison, see our agent framework comparison guide.

11Why Lushbinary for AI Agent Development

At Lushbinary, we've been building production AI agents on AWS since the early days of Bedrock. We help teams:

• Design agent architectures that balance cost, latency, and reliability
• Implement MCP Gateway integrations connecting agents to internal tools and data
• Set up observability and guardrails for safe autonomous operation
• Migrate from self-hosted frameworks (LangChain, CrewAI) to managed AgentCore
• Build custom coding agents for CI/CD, code review, and incident response

❓ Frequently Asked Questions

📚 Sources

• Amazon Bedrock AgentCore — Official Product Page
• AgentCore Pricing — AWS
• Get Started with the AgentCore CLI — AWS Documentation
• AgentCore Getting Started Guide — AWS

Content was rephrased for compliance with licensing restrictions. Pricing and feature details sourced from official AWS documentation as of April 2026. Pricing may change — always verify on the AWS pricing page.