Building AI Agents on AWS with Amazon Bedrock AgentCore

2026 has been the year AI agents moved from demo to production. Writing a chatbot that calls a model was never the hard part. The hard part is the plumbing around it: the orchestration loop, tool execution, memory, identity, session isolation, observability and governance. AWS’s answer is Amazon Bedrock AgentCore, and it has had a busy year. A managed agent harness went generally available in June 2026, and a next-generation Runtime followed in September. This guide covers what changed, what it means for engineering and business leaders, and how to take an agent from prototype to production on AWS without making the usual mistakes.

What Is Amazon Bedrock AgentCore?

AgentCore is AWS’s platform to “build, connect, and optimize agents” (AWS Machine Learning Blog). It doesn’t lock you into one framework. AWS says it works with frameworks teams already use, including LangGraph, LlamaIndex, CrewAI and Strands Agents (AWS). Think of it as the managed operational layer under your agents: compute, tool gateway, memory, identity and tracing, with security controls built in from the first call.

What Changed in 2026

1. The managed agent harness (GA June 2026)

AWS describes it this way: “If the model is the brain, the harness is the body.” The harness runs the orchestration loop, executes tools, manages the context window, persists state across turns, recovers from failures and isolates each session (AWS What’s New, June 2026). You don’t write that loop yourself. You define an agent in configuration (model, tools, skills and instructions), and AWS says you can run it in just three API calls (AWS).

Two details matter for architects:

• Model independence. The harness is decoupled from the model, so you can switch providers mid-session without losing context. For example, you can plan with one model and write code with another.
• No dead end. When you need custom orchestration, a single CLI command exports the harness to Strands-based code on the same platform.

When the harness first arrived in preview in April, each session got its own microVM with filesystem and shell access. Filesystem persistence let agents suspend mid-task and resume exactly where they left off (AWS What’s New, April 2026).

2. AgentCore CLI and infrastructure as code

The AgentCore CLI keeps prototyping, deployment and operations in one terminal workflow. Deployments run through infrastructure as code: AWS CDK is supported now, and AWS has said Terraform is “coming soon” (AWS). Your agent configuration becomes version-controlled and reproducible, the same discipline you apply to the rest of your cloud estate.

3. The new AgentCore Runtime (September 2026)

On 18 September 2026, AWS announced the next generation of AgentCore Runtime, its serverless microVM compute for agents. Memory is now elastic: unused memory is reclaimed during the session, so you pay for actual usage rather than peak. Cold starts restore from a snapshot. In AWS’s testing, P75 cold starts were 1.9–2.0 seconds for container images from 200 MB to 2 GB, compared with 5.4–30 seconds on V1. Scale to zero and hardware-enforced session isolation are still included. To opt in, you set platformVersion to V2 (AWS What’s New, Sept 2026).

4. Knowledge, evaluation and guardrails

AWS also made several features generally available: Bedrock Managed Knowledge Base, Web Search, Guardrail integration, and recommendations and A/B testing. Insights and agent payments are in preview. Recommendations analyze traces and evaluation outputs and suggest changes to system prompts and tool descriptions. A/B testing splits live production traffic between agent versions. AWS says these work wherever your agents run: on AgentCore, AWS Lambda, Amazon EKS, or outside AWS (AWS Machine Learning Blog).

A Practical Roadmap: Prototype to Production

Step 1: Pick a narrow, measurable use case

Start with a task that has clear inputs, a small toolset and an obvious success metric. Good examples are triaging support tickets, answering questions from internal policy documents, or summarizing cloud cost anomalies for FinOps teams. Avoid “do everything” assistants in your first release.

Step 2: Prototype with the harness

Define the model, the system prompt and two or three tools in configuration. Because you can override settings per invocation without redeploying, test several models and prompt variants side by side before you commit (AWS).

Step 3: Build an evaluation dataset early

Collect 50–100 real requests with expected outcomes. Use batch evaluation to score each change, so improvements are measured rather than guessed. Regressions get caught before they reach users.

Step 4: Lock down identity and tool access

Treat every tool an agent can call as a privileged API. Route tools through the gateway that enforces security policies, apply least-privilege IAM, attach Guardrails, and log every action. If your industry is regulated, align agent controls with your existing compliance framework. Our security, governance and compliance team can help map these controls.

Step 5: Deploy as code and optimize cost

Promote the validated agent with the CLI and CDK, and choose Runtime V2 where it’s available. At the September launch, V2 was in us-east-1, us-east-2, us-west-2, eu-west-1 and ap-northeast-1 (AWS). Check the region list against your data residency needs. This matters especially for teams in India and the UAE.

Step 6: Improve continuously

Once the agent is live, use traces, recommendations and A/B testing to refine prompts and tool descriptions against real traffic.

Common Pitfalls to Avoid

• Skipping evaluation: without a test set, every prompt change is a gamble.
• Over-permissioned tools: an agent with broad write access is a security incident waiting to happen.
• Ignoring regions: a feature that is GA in one region may be missing in the region your data must stay in.
• Hand-built plumbing: custom orchestration code you have to maintain forever, when a managed harness would do.

Frequently Asked Questions

Is AgentCore only for agents built on Amazon Bedrock models?

No. The harness is model-agnostic and lets you switch models mid-session. AgentCore also works with open-source frameworks such as LangGraph, CrewAI, LlamaIndex and Strands Agents.

Do I need to write orchestration code?

Not to get started. The managed harness runs an agent defined entirely in configuration. If you outgrow it, you can export to Strands-based code.

How is AgentCore Runtime billed?

It is serverless and usage-based, with scale to zero. With the new Runtime, unused memory is reclaimed during the session, so you pay for actual usage rather than peak. See the AWS AgentCore pricing page for current rates.

Can I evaluate agents running outside AgentCore?

Yes. AWS says recommendations and A/B testing work for agents on AgentCore, AWS Lambda, Amazon EKS, or non-AWS environments.

Ready to Put AI Agents to Work?

Shanav is an AWS Advanced Tier Partner with offices in India and the UAE. We help teams design, secure and run production-grade generative AI on AWS. To see how we approach it, look at our generative AI and innovation services. When you’re ready to plan your first agent, talk to our AWS experts.