Getting an agent working has all the time meant fixing a protracted record of infrastructure issues earlier than you may check whether or not the agent itself is any good. You wire up frameworks, storage, authentication, and deployment pipelines, and by the point your agent handles its first actual activity, you’ve spent days on infrastructure as a substitute of agent logic.
We constructed AgentCore from the bottom as much as assist builders give attention to constructing agent logic as a substitute of backend plumbing, working with frameworks and fashions they already use, together with LangGraph, LlamaIndex, CrewAI, Strands Brokers, and extra. Right this moment, we’re introducing new capabilities that additional streamline the agent constructing expertise, eradicating the infrastructure obstacles that gradual groups down at each stage of agent improvement from the primary prototype by manufacturing deployment.
Go from thought to a working agent in three steps
Each agent has an orchestration layer which accommodates the loop that calls the mannequin, decides which software to invoke, passes outcomes again, manages context home windows, and handles failures. Working that loop requires infrastructure beneath it: compute to host the agent, a sandbox to securely execute code, safe connections to instruments, persistent storage, and error restoration. This infrastructure kinds the agent harness, enabling an agent to truly run.
Till now, constructing that harness was the very first thing each group needed to do from scratch. That meant selecting a framework, writing the orchestration code, connecting it to instruments and reminiscence, and guaranteeing authentication, all earlier than the agent may course of a single request. It’s essential work, but it surely’s not the work that tells you whether or not your agent goes to be helpful. Most groups we’ve labored with spent days on this infrastructure earlier than they may run their first actual check.
The brand new managed agent harness characteristic in AgentCore replaces all that upfront construct with a simple configuration. You declare your agent and run it in simply three API calls, with out writing orchestration code. You outline what your agent does: which mannequin it makes use of, which instruments it might probably name, and what directions it follows. AgentCore’s harness stitches collectively compute, tooling, reminiscence, identification, and safety to create a working agent you could check in minutes. Attempting a unique mannequin or including a software is a config change, not a code rewrite. You’ll be able to check a number of variations of an agent in minutes by altering the API parameter on the fly.
That velocity doesn’t come at the price of flexibility. The harness in AgentCore is powered by Strands Brokers, the open supply framework from AWS. If you want customized orchestration logic, specialised routing, or multi-agent coordination, you turn from config to code-defined harness, with the identical platform, similar microVM isolation, similar deployment pipeline. AgentCore persists session state to a sturdy filesystem, so brokers can droop mid-task and resume precisely the place they left off. This makes human-in-the-loop patterns sensible with out customized plumbing, and with out redesigning the agent later when these wants come up. You will get began in minutes then add extra capabilities and management when your wants evolve, with none rearchitecture.
“We’re constructing AI brokers that can revolutionize ecommerce”, stated Rodrigo Moreira, VP of Engineering, VTEX. “Beforehand, prototyping every new agent required days of orchestration code and infrastructure setup earlier than we may validate an thought. The harness characteristic in AgentCore will change that: swapping a mannequin, including a software, or refining an agent’s directions is now a configuration change, not a rebuild. We will now validate agent concepts in minutes as a substitute of days, and we’re trying ahead to accelerating agent improvement additional with these new capabilities”.
Construct, deploy, function your brokers from the identical terminal
You’ve received your agent working, and now you wish to run it in manufacturing. That normally means stepping out of your editor, organising a deployment pipeline, configuring environments, and stitching collectively a course of that appears nothing just like the workflow you used to construct the agent within the first place.
The brand new AgentCore CLI retains you in a single workflow throughout the total lifecycle: prototype, deploy, function, from the identical terminal that you just’re already working in. You iterate in your agent regionally, and when it’s prepared, you deploy it with out switching instruments or constructing a separate pipeline. AgentCore powers deployments by infrastructure as code (IaC) with CDK help and Terraform (coming quickly), so your agent configuration is reproducible and version-controlled. What you examined regionally is strictly what runs in manufacturing.
Give your coding brokers the suitable context
All through the agent improvement journey, most builders are working alongside a coding assistant, akin to Claude Code or Kiro. However a coding assistant is barely as efficient because the context it has. A general-purpose MCP server may give it entry to APIs and documentation, but it surely doesn’t encode the opinions that matter: which patterns to make use of, how capabilities match collectively, what the beneficial path appears to be like like for frequent duties. New pre-built abilities in AgentCore transcend uncooked API entry. They offer coding brokers curated, present information of AgentCore greatest practices, so the options you get replicate how the platform is supposed for use, not solely what endpoints exist. Kiro already consists of this right now as a built-in Energy, with Plugins for Claude Code, Codex, and Cursor coming quickly. On a platform that evolves rapidly, having correct context in your coding agent means fewer fallacious turns from the very first line of code.
Get began
The managed agent harness in AgentCore is accessible in preview right now in 4 AWS Regions: US West (Oregon), US East (N. Virginia), Asia Pacific (Sydney), and Europe (Frankfurt). AgentCore CLI and protracted agent filesystem, can be found AWS business Regions the place AgentCore is obtainable. Coding agent abilities can be accessible by the tip of April. You pay just for the assets that you utilize, with no extra cost for the CLI, harness, or ability (be taught extra in AgentCore pricing web page). Go to AgentCore Documentation to get began.
You should utilize these capabilities to remain centered on agent logic, with out worrying in regards to the infrastructure setup. As your agent evolves, you add evaluations, reminiscence, software connections, and coverage enforcement with out rearchitecting. The platform that you just prototype on is similar one you run in manufacturing.
In regards to the authors
Madhu Parthasarathy
Madhu Parthasarathy is the GM of Amazon Bedrock AgentCore, with over 20 years of experience in constructing giant scale distributed infrastructure. Madhu has been with Amazon for over 16 years, the place he led a number of initiatives in Amazon Retail, Elastic Block Retailer, and extra lately, AgentCore. Madhu has held varied management positions at different corporations together with LinkedIn, the place he led Enterprise platform that powered all LinkedIn enterprise traces of companies and a neo-cloud startup, the place he led AI infrastructure, driving the imaginative and prescient for safety and developer expertise. Madhu is presently based mostly in Santa Clara, California.
