AWS Step Functions Workflows Are Now Tools Your AI Agents Can Call

AWS Step Functions Workflows Are Now Tools Your AI Agents Can Call

AWS made a quiet change last week that matters more than most model launches. Step Functions, which already orchestrates over 220 AWS services, can now be handed to an agent as a tool. Workflows your team already trusts became callable without a rewrite, through Amazon Bedrock AgentCore Gateway, which connects agents to tools over the Model Context Protocol.

The idea is simple and powerful. Teams have spent years building reliable workflows in Step Functions: order processing, data pipelines, approval chains, incident response. Those workflows have retries, error handling, timeouts, and audit trails built in. Until now, an AI agent could not call them directly. Now it can, and the workflow runs exactly as it would if a human triggered it, with the same guardrails.

That flips the usual agent problem around. Instead of asking a model to figure out how to do a multi-step task and hoping it does not skip a step, you give the agent a tool that already knows the steps. The agent decides when to call it and what to pass in. The workflow decides how the work gets done. It is the difference between handing someone a recipe and handing them a kitchen that already knows the recipe.

The rest of the AWS week

The Step Functions change headlined the AWS Weekly Roundup for November 2, but it was not the only launch. OpenAI's GPT-6.1 Terra is now generally available on Amazon Bedrock, priced below the model beside it, according to OpenAI's figures rather than an independent test. For teams already running on Bedrock, a cheaper capable model in the same console is a straightforward win.

Two post-quantum launches landed for a deadline nobody can put on a calendar yet. AWS Key Management Service now supports ML-KEM for post-quantum key encapsulation, with private keys in FIPS 140-3 Level 3 validated hardware security modules. AWS IAM Identity Center now supports cross-Region trusted identity propagation, so you can carry a user's identity with their request to AWS services in another Region.

On the infrastructure side, Amazon ECS now supports GPU sharing with NVIDIA Multi-Process Service, so you can run multiple ECS tasks concurrently on a single NVIDIA GPU. That is the practical answer to work that only ever uses a fraction of one GPU, like small-model inference and multi-model serving. AWS Lambda supports Seekable OCI lazy loading for Lambda Managed Instances, and Amazon EKS Auto Mode now supports the Kubernetes Gateway API, so you can declare how traffic reaches your applications with Gateway and HTTPRoute resources without installing a separate controller.

The Nubank example

The roundup also covered a real-world case worth reading. Nubank, which serves more than 140 million customers across Brazil, Mexico, and Colombia, automated disaster recovery with the AWS DevOps Agent. The bank classifies over 300 business flows as critical, each tested on a regulator's schedule. Automating that testing is the kind of unglamorous work that decides whether a cloud migration actually pays off, and it is the same kind of work Step Functions as agent tools is aimed at: the reliable, repeatable processes that already exist.

What this means for builders

If you are building agents on AWS, the practical move is to inventory the Step Functions workflows you already have and ask which ones an agent should be allowed to call. Start with read-only or low-risk workflows, like status checks and report generation, before handing an agent the one that moves money or deletes data. AgentCore Gateway handles the connection over MCP, but the access control is still your job.

If you are not building agents yet, the launch is still useful, because it lowers the cost of starting. You do not need to rebuild your business logic for an agent world. You need to expose the workflows you already trust, one at a time, and let the agent call them. The AWS European Sovereign Cloud test showed how seriously AWS takes isolation and control, and the SoftBank GPU Cloud launch showed how fast AI infrastructure is expanding. This launch is the software half of the same story: the cloud is becoming something agents use, not just something humans use.

AWS re:Invent returns to Las Vegas from November 30 to December 4, with more than 2,200 sessions. Expect AgentCore and agent tooling to be a major theme. The question for re:Invent is not whether agents will use your workflows, but which of your workflows you will let them touch first.