Run production AI agents in n8n with Amazon Bedrock AgentCore harness
AWS Sundar Raghavan ● Covered by 3 sources
AWS shipped an open-source n8n node for Bedrock AgentCore, its production agent harness. Now n8n bots get memory, tools, and VPC privacy without extra agent code.
Based on reporting by AWS, Sundar Raghavan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
n8n's built-in AI Agent node has always been fine for bolting a model call onto a workflow. But anyone who's tried to push it toward something real quickly hits the ceiling: no memory that survives past one run, no way to hand the model an actual tool like a code sandbox, no room for a task that takes more than one turn. AWS is closing that gap with a new open-source community node, @aws/n8n-nodes-agentcore, that plugs n8n directly into Amazon Bedrock AgentCore harness, now generally available. The harness itself is built on Strands Agents, AWS's open-source agent framework, and the node ships under the MIT license.
The design is deliberately simple: one operation, one deciding field called Harness ARN. Leave it blank and the node builds an agent on first run, then reuses and updates it as your configuration changes. Paste in an existing harness ARN and the node just invokes that agent, letting you point n8n at something you built elsewhere. Credentials follow the same pattern n8n already uses for its AWS Lambda and S3 nodes, so anyone automating AWS today won't find the setup unfamiliar. And you're not locked to Bedrock models either — the node also talks to OpenAI, Google Gemini, and LiteLLM-supported providers, with the option to switch providers between turns of the same conversation.
Memory is on by default and scoped in a hierarchy: the agent holds shared configuration, an Actor ID isolates one person's history from another's, and a Session ID isolates individual conversations within that actor. In a walkthrough example, a travel-concierge agent remembers a user's preference for warm beaches and vegetarian food across turns just by reusing the same session ID, with input token counts rising on later turns as it loads the prior conversation.
Tools are where things get more interesting. A code interpreter tool lets the agent actually run code in a sandbox rather than guess — one example has it generating 500 random exam scores and computing the real mean, median, and standard deviation instead of estimating them. Skills, pulled from AWS's curated catalog, a Git repo, S3, or a filesystem path, give the agent domain knowledge loaded only when a task calls for it. For teams that need private networking, the credential itself can be switched to VPC mode, and because the harness pulls its container image from a private ECR repo in the same region, subnets need endpoints for ECR and S3 rather than a NAT gateway.
None of this is free to leave running. The harness, its managed memory store, and any VPC endpoints are billed AWS resources, and AWS points users toward CLI commands to list and delete harnesses once a walkthrough or project is done.
My take — AI-written commentary, not fact-checked reporting
This is AWS doing what AWS does best: taking something genuinely useful and open-source-flavored and quietly wiring it back into its own billed infrastructure, right down to reusing the Lambda and S3 credential pattern so it feels native. Multi-provider support is a real olive branch to people who don't want to be locked into Bedrock models, but the harness itself, the memory store, and the VPC endpoints all still run on AWS's meter. No-code builders should read the cleanup section before the memory-and-tools demo, because
Read more about this at: AWS