AWS now lets an AI agent start your existing Step Functions workflows as tools, without letting it change which steps run. Per the AWS announcement on 29 October, the agent decides when to run a workflow and what input to pass. Step Functions then runs each step exactly as defined. For teams that do not trust a model to follow refund policy alone, that split is the point.
What shipped
Step Functions runs a process as a fixed sequence of steps, called a state machine. The new piece is a built-in connector in Amazon Bedrock AgentCore Gateway. That service hands tools to agents over the Model Context Protocol (MCP), an open standard for connecting agents to tools. The AgentCore connector documentation says you pick which of ten operations become tools, with no Lambda function, MCP server or hand-written schemas.
The control that matters is pinning. You fix the state machine on the start and list operations during setup. The agent never sees that setting, and the gateway rejects calls that try to override it. A support agent can start the refund workflow, but it cannot aim the same tool at payroll.
What it costs
There is no separate fee; AWS says standard rates apply. The AgentCore pricing page lists gateway calls at $0.005 per 1,000. The Step Functions pricing page lists Standard workflows at $0.000025 per state transition in US East (N. Virginia), after 4,000 free a month.
With made-up numbers: 10,000 agent-started runs of a workflow with six state transitions, each with one start call and one status check, means 20,000 gateway calls ($0.10) and 56,000 billable transitions ($1.40). Tool-list calls, model calls and agent hosting come on top.
Setup time
The AWS tutorial runs to six steps, from creating a state machine and an IAM role to test calls and cleanup. Creating the gateway and the connector can each take a few minutes. It assumes the AWS CLI and signed requests, so give it to someone who knows IAM. We have not run the tutorial ourselves.
Where it breaks
Most sharp edges come from the two workflow types. Express workflows, built for high-volume work, run for at most five minutes, per the Step Functions guide to workflow types. An agent gets an Express result back in one call but cannot look the run up later. Pausing a workflow until the agent finishes a task works only on Standard.
- One pinned state machine per operation per connection, so five workflows means five connections.
- All connections on a gateway share one IAM role. For different permissions, AWS says use separate gateways.
- Skip the pin and the connection still gets created, but its tool fails on every call.
- Status, history and stop calls are not pinned. Only the IAM role limits which runs they reach, so scope it to your workflows.
- The calls that resume a paused workflow cannot be scoped in IAM. The task token is the only lock, so treat tokens as secrets.
- CloudTrail, the AWS audit log, shows the gateway’s role as the caller. Log which user’s agent started a run yourself.
Who it is for
Teams already on Step Functions get a low-cost way to let an agent trigger approved processes or report failed runs. The gateway only reaches state machines in its own AWS Region, so both must sit in the same Region. The connector does not make the agent better at deciding when to start a workflow; that still sits with your model and prompt.
What to do this week
Pick one Standard workflow you already trust, such as a refund. Write down who may start it and with what input. Then run the AWS tutorial in a test account with a role scoped to that one state machine. Check the agent’s tool list before connecting anything that moves money.
Sources
Get the weekly email
The week's most useful AI news, a tutorial and an interview.
Anna Jakovleva
Anna Jakovleva is the senior editor for AI tools at AI Observer, leading coverage of agents and automation, design and product, and the site's tutorials. She worked in operations and product teams before joining, so she judges tools by setup time, cost and where they break. Her writing is hands-on and practical.
Spotted a mistake? Tell us on the corrections page.