Short answer: On 29 October 2026, AWS made Step Functions workflows available as AI agent tools through Amazon Bedrock AgentCore Gateway. A built-in connector exposes 10 Step Functions operations as MCP tools, so an agent picks when to run a workflow and what input to pass, while the state machine fixes the steps and their order. There is no connector fee: Standard workflows cost $0.025 per 1,000 state transitions (4,000 free a month) and Gateway tool calls cost $0.005 per 1,000. By our arithmetic, 100,000 agent-started runs of a 6-step workflow cost about $16 a month before model and Lambda costs.
Facts from the AWS announcement, AgentCore and Step Functions documentation, and both pricing pages (US East), checked 1 November 2026. We haven’t deployed the connector ourselves.
Most AI agent failures in business processes come from the same place: the model is asked to both decide and execute. It skips a step, calls tools in the wrong order, or retries a payment it shouldn’t. AWS’s 29 October announcement splits those jobs. The agent decides, and a Step Functions state machine executes every step “exactly as you defined it”. For B2B teams already running processes on AWS, this is one of the cleaner ways yet to put a deterministic workflow behind an AI agent.
What AWS launched
According to the AWS What’s New post, you can now use existing Step Functions Standard and Express workflows as tools for AI agents through Amazon Bedrock AgentCore Gateway, which connects agents to tools over the Model Context Protocol (MCP). AWS’s example: a support agent starts a refund workflow, and Step Functions runs the eligibility check, manager approval and credit steps, “so every refund follows your policy”. An operations agent can also look up which runs failed overnight and where.
The connector documentation adds the details that matter for builders:
- No glue code. You add a gateway target with
connectorId: "states". The gateway generates the tool schemas and authenticates to Step Functions with its service role. You don’t write a Lambda function or an MCP server. - No changes to existing state machines. Standard and Express workflows can be exposed as they are.
- Pinned workflows. You pin a
stateMachineArnto each operation that takes one. The agent never sees it, and if it tries to pass one, the gateway rejects the call. - Availability: every AWS Region where AgentCore Gateway is available. The gateway calls Step Functions in its own Region, so the state machine must be in the same Region.
Which operations can an agent call?
The connector exposes ten Step Functions operations as MCP tools. Each tool name is prefixed with the target name, for example orders___StartExecution.
| Operation | Who sets the target | Works with | What the agent gets |
|---|---|---|---|
| StartExecution | Pinned by you | Standard and Express | Execution ARN and start date |
| StartSyncExecution | Pinned by you | Express only | Status, output and timing in the same call |
| ListExecutions | Pinned by you | Standard only | Recent runs, filterable by status |
| DescribeStateMachine | Pinned by you | Both | Metadata, or the definition with ALL_DATA |
| DescribeExecution | Agent passes execution ARN | Standard executions only | Status, plus input and output with ALL_DATA |
| GetExecutionHistory | Agent passes execution ARN | Standard executions only | Event list for each step |
| StopExecution | Agent passes execution ARN | Standard executions only | Stop confirmation |
| SendTaskSuccess / SendTaskFailure / SendTaskHeartbeat | Agent passes task token | Both (callback pattern needs Standard) | Resumes or fails a paused task |
By default, each tool shows the agent only a minimal set of fields. StartExecution, for example, advertises just input, a JSON string. You can expose a few extra fields, such as name or maxResults, through parameter overrides.
Standard or Express: which workflow type fits an agent?
The Step Functions documentation sets very different limits for the two workflow types, and they decide which tools work.
| Standard | Express | |
|---|---|---|
| Maximum duration | 1 year | 5 minutes |
| Execution semantics | Exactly once | At least once (async), at most once (sync) |
| Execution history | Via API for 90 days after completion | Only through CloudWatch Logs, if enabled |
| Callback (.waitForTaskToken) | Supported | Not supported |
| Price (US East) | $0.000025 per state transition; 4,000 free a month | $1.00 per million requests + $0.00001667 per GB-second |
| Best agent use | Payments, approvals, onboarding, anything auditable | Fast lookups where the agent needs the result now |
The practical rule: if the action must not run twice, such as issuing a refund or creating an invoice, use a Standard workflow. Its exactly-once model and 90-day history give you an audit trail of every step the agent triggered. If the agent needs an answer within the same turn, use an Express workflow with StartSyncExecution. Note that the StartSyncExecution API reference says it returns HTTP 200 even when the execution fails, so the agent has to read the returned status.
What does it cost?
AWS says you “pay standard rates for Step Functions and AgentCore Gateway”, and neither pricing page lists a separate connector fee. The relevant rates from the Step Functions pricing page and AgentCore pricing page:
| Item | Price (US East) |
|---|---|
| Standard workflow state transitions | $0.000025 each ($0.025 per 1,000), first 4,000 a month free, free tier doesn’t expire |
| Express workflow requests | $1.00 per million |
| Express workflow duration | $0.00001667 per GB-second ($0.06 per GB-hour for the first 1,000 GB-hours), billed in 64 MB memory chunks |
| Gateway API invocations (ListTools, InvokeTool, Ping) | $0.005 per 1,000 |
| Gateway Search API | $0.025 per 1,000 |
| Gateway tool indexing | $0.02 per 100 tools indexed per month |
| AgentCore Identity used through Gateway | No additional charge |
| AgentCore Policy authorization (optional) | $0.000025 per request |
Retries count as extra state transitions, and any Lambda functions or other AWS services your workflow calls are billed at their own rates.
Here’s what that means for a simple agent tool. We assume a Standard workflow with 6 state transitions per run and 3 Gateway calls per run (list tools, start the execution, check its status):
| Agent-started runs per month | State transitions cost | Gateway cost | Total |
|---|---|---|---|
| 10,000 | $1.40 | $0.15 | ~$1.55 |
| 100,000 | $14.90 | $1.50 | ~$16.40 |
| 1,000,000 | $149.90 | $15.00 | ~$164.90 |
Our arithmetic from AWS list prices (US East), after the 4,000 free state transitions. Excludes model inference, Lambda, other AWS services, data transfer and the agent runtime, which usually cost more than the orchestration itself.
Why this matters for B2B automation
- Deterministic workflows are becoming the standard agent tool. The agent chooses, the workflow executes. ServiceNow made a similar move this week by exposing its workflows to ChatGPT, which we covered in ServiceNow and OpenAI. Expect more vendors to follow.
- Governance moves into the platform. Pinned state machines, IAM-scoped service roles and Standard execution history give security teams something they can review, instead of trusting a prompt.
- Human-in-the-loop gets a clear pattern. A Standard workflow can pause on a callback task and wait for an agent, or a human working through an agent, to send
SendTaskSuccess. That is a cleaner approval step than asking a model to “wait for confirmation”.
How does this compare with exposing n8n workflows to agents?
n8n offers the same pattern without AWS. Its MCP Server Trigger node publishes n8n workflows as MCP tools over SSE or streamable HTTP, with bearer or header authentication, and the MCP Client Tool lets n8n’s own agents call external MCP servers.
| Step Functions via AgentCore Gateway | n8n MCP Server Trigger | |
|---|---|---|
| Hosting | AWS-managed | n8n Cloud or self-hosted |
| Pricing basis | Per state transition or request + per Gateway call | Your n8n plan, or your own server if self-hosted |
| Agent authentication | AWS IAM (SigV4) in AWS’s tutorial | None, bearer or header auth |
| Workflow can’t be changed by the agent | Yes, state machine pinned | Yes, the agent can only call attached tools |
| Long-running, exactly-once runs | Standard workflows, up to 1 year | Depends on your workflow design and n8n setup |
| Best for | Teams already on AWS with IAM and audit needs | SaaS-heavy stacks, agencies, self-hosters |
If your CRM, email and billing automations already run in n8n, the MCP Server Trigger is the shorter path. Read our guides to securing your n8n instance and n8n queue mode first: n8n’s docs say MCP connections need all /mcp* traffic routed to a single webhook replica when you run several.
Before you hand any workflow to an agent, scan it with the free n8n Workflow Auditor, estimate execution volumes in the n8n Cost & Hosting Planner, and compare the AI subscriptions in your stack with the AI SaaS Pricing Comparison.
How to try it
AWS’s Step Functions tutorial walks through the setup in six steps:
- Create a Standard state machine (the tutorial uses a single Pass state).
- Create a gateway service role whose IAM policy is scoped to that state machine and its executions.
- Create a gateway with
--protocol-type MCPand--authorizer-type AWS_IAM, then add a target withconnectorId: "states"and the operations you want to expose. - Call
tools/listandtools/callwith a SigV4-signed client. The tutorial usesawscurl. - Confirm the run with
DescribeExecutionandGetExecutionHistory. - Delete the gateway target, gateway, state machine and roles to avoid ongoing charges.
All targets on a gateway share one service role, so AWS recommends a separate gateway for each workflow that needs a different IAM boundary.
What we couldn’t verify
- Real-world latency of a tool call through the gateway to
StartSyncExecution. AWS doesn’t publish it, and we didn’t deploy the connector. - The exact list of Regions. AWS says “all AWS Regions where AgentCore Gateway is available” but the announcement doesn’t name them.
- Prices outside US East. All rates above are AWS’s US East (N. Virginia) examples; other Regions can differ.
- How many Gateway calls a real agent makes per run. Our 3-call assumption is illustrative; agents that poll often or list tools every turn will make more.
Frequently asked questions
What did AWS announce on 29 October 2026?
AWS made Step Functions workflows available as tools for AI agents through Amazon Bedrock AgentCore Gateway. A built-in Step Functions connector turns existing Standard and Express state machines into Model Context Protocol (MCP) tools, so an agent decides when to run a workflow and what input to pass, and Step Functions runs the steps exactly as defined. No Lambda function or custom MCP server is needed.
How much does it cost to let an agent run Step Functions workflows?
AWS says you pay standard rates for both services, with no extra fee for the connector. Standard workflows cost $0.000025 per state transition ($0.025 per 1,000) after 4,000 free transitions a month. Express workflows cost $1.00 per million requests plus duration. AgentCore Gateway charges $0.005 per 1,000 tool invocations such as ListTools and InvokeTool. Model inference is billed separately.
Which Step Functions operations can an agent use?
The connector exposes ten operations. StartExecution, StartSyncExecution, ListExecutions and DescribeStateMachine are pinned to a state machine you choose, and the agent can’t change it. DescribeExecution, StopExecution and GetExecutionHistory take an execution ARN, and SendTaskSuccess, SendTaskFailure and SendTaskHeartbeat take a task token.
Should I use Standard or Express workflows for agent tools?
Use Standard for long-running, auditable or non-idempotent work such as payments or approvals: it runs up to one year, executes exactly once and keeps history for 90 days. Use Express with StartSyncExecution when the agent needs the answer in the same tool call: it runs up to five minutes and returns status and output synchronously, but its executions can’t be described or stopped by ARN afterwards.
Can agents pause a workflow for human or agent approval?
Yes, on Standard workflows. A task that uses the .waitForTaskToken callback pattern pauses the workflow and sends a task token to the agent’s application. The agent resumes it with SendTaskSuccess or SendTaskFailure. IAM can’t limit those calls to specific executions, so AWS says to treat task tokens as secrets.
Can I do the same thing with n8n?
Yes. n8n’s MCP Server Trigger node exposes n8n workflows as MCP tools over SSE or streamable HTTP, with bearer or header authentication. It’s a good fit if your automations already live in n8n or you self-host. Step Functions via AgentCore Gateway fits teams whose processes already run on AWS and need IAM controls and exactly-once Standard executions.
Related reading
- Box Architect: AI that designs agentic workflows (BoxWorks 2026)
- ServiceNow and OpenAI: ServiceNow workflows now run from ChatGPT
- New AI agent tools this week, checked (30 October 2026)
- Zapier vs Make vs n8n AI agents
- Salesforce Agent Fabric and agent evals
- Stripe Treasury for Agents
Sources
AWS: Step Functions workflows now available as AI agent tools (29 October 2026); AgentCore Gateway: Step Functions as connector target; Step Functions tutorial: use workflows as tools for AI agents; Step Functions: choosing a workflow type; StartSyncExecution API reference; Step Functions pricing; AgentCore pricing; n8n docs for the MCP Server Trigger and MCP Client Tool. All pages checked 1 November 2026. Cost tables are our own arithmetic.
