Step Functions workflows can now be called by AI agents as tools

Amazon Web Services says existing Step Functions workflows can now be used as tools by AI agents through Amazon Bedrock AgentCore Gateway. The agent chooses when to start a workflow; the service runs every step as defined.

Agents call workflows, workflows keep the order

According to AWS, both Standard and Express workflows qualify, so teams do not need to rebuild anything. AgentCore Gateway links agents to tools using the Model Context Protocol (MCP), and Step Functions workflows now appear there as one more kind of tool. In AWS’s words: “Your agent decides when to run a workflow and what input to pass.” The workflow itself then proceeds without the model improvising.

AWS gives two examples. A customer support agent could start a refund workflow, where Step Functions performs the eligibility check, the manager approval and the credit, so each refund follows company policy. An operations agent could also look up past runs and tell an on-call engineer which order runs failed overnight and at which step.

Step Functions is AWS’s visual workflow service, which the company says can orchestrate more than 220 AWS services. The feature is available in every AWS Region where AgentCore Gateway is offered. There is no separate charge: customers pay the usual rates for Step Functions and for AgentCore Gateway. AWS points to a tutorial in the Step Functions Developer Guide and to a page on Step Functions connector targets in the AgentCore Gateway documentation. It also mentions the Agent Toolkit for AWS, which gives coding assistants such as Kiro guidance for building on AWS.

Deterministic steps for non-deterministic agents

The practical change is where control sits. An agent that has to carry out a multi-step process through free-form tool calls can skip a check or reorder steps. Putting the process inside a workflow moves that sequencing out of the model, leaving the agent to decide only whether to start the process and with what input. Teams that already run order, billing or approval flows on Step Functions can expose them without writing a custom MCP server for each one, though AWS does not say how much setup work remains in practice.

The source does not say how agents are authorized to start particular workflows, how failures are returned to the model, or how latency compares with calling a workflow directly. Those details sit in the documentation AWS links, and AWS’s claims about consistency have not been independently verified. For readers new to the protocol, our MCP explainer covers the basics.

AgentCore Gateway becomes the single front door

The launch fits a pattern of AWS positioning AgentCore Gateway as the place where an enterprise’s existing systems become agent tools, rather than asking customers to adopt a new runtime for each. An AWS Machine Learning blog post published the same day describes how the marketing platform MoEngage used the gateway to turn about 400 internal REST APIs, built by 25 engineering teams, into tools for 10 internal agents across seven AWS Regions. According to that post, the gateway let the company avoid building more than 20 separate MCP servers, and it estimates the saving at about six months of platform engineering effort.

The same post points to a design concern that matters for any large tool catalog: listing every tool would cost roughly 50 tokens each, or more than 20,000 tokens per listing, so MoEngage relies on semantic search and Cedar-based policies that hide restricted tools from each caller. Those figures come from the customer’s own account, not an outside test. Still, they suggest why AWS is extending the gateway to more tool types: the more kinds of capability sit behind one MCP endpoint, the more an enterprise’s agent traffic, and its billing for both gateway and underlying services, runs through AWS.

Step Functions adds an orchestration layer to that bet. APIs are single calls; workflows are whole business processes with approvals and retries. Offering them through the same interface lets agents handle longer jobs while the company’s own rules stay in force. Whether customers trust agents to start sensitive processes such as refunds will depend on the permission controls, which the announcement does not describe.