Orbital / AWS Step Functions as AI agent tools: a business control checklist

AWS Step Functions as AI agent tools: a business control checklist.

AWS now lets AI agents invoke Step Functions workflows through AgentCore Gateway. Evaluate workflow scope, IAM, task tokens, approvals, replay safety, logs and full cost.

AI agents · Connected workflows

By Orbital Content Team · Published November 3, 2026

AI-assisted research and drafting with editorial review. Release date, availability, connector operations, IAM, authentication, observability, execution-history and pricing statements were checked against official AWS sources on November 3, 2026. AWS product descriptions are identified as vendor-reported. The evaluation framework, authority contract, control matrix, risk register, cost model and pilot are Orbital analysis. Orbital did not configure or test AgentCore Gateway, invoke a Step Functions workflow, inspect an AWS account, reproduce AWS claims, audit IAM or logs, measure performance or cost, or review customer-specific contracts, architecture or compliance requirements.

AWS announced on October 29, 2026 that existing AWS Step Functions Standard and Express workflows can be exposed as tools for AI agents through Amazon Bedrock AgentCore Gateway. The gateway presents configured workflow operations over the Model Context Protocol, creates the tool schemas and authenticates to Step Functions with a gateway service role.

The useful idea is a division of responsibility: an agent decides when to call a tool and what approved input to provide, while the state machine controls the ordered workflow. That can keep deterministic validation, approval and execution steps outside the model. It does not make the trigger, input, permissions, data, downstream effects or recovery safe automatically.

What AWS released on October 29

AWS says AgentCore Gateway now includes a built-in Step Functions connector. A team can select supported operations, point the target at an existing state machine where required and let a connected agent discover the resulting MCP tools. AWS says the capability is available in every region where AgentCore Gateway is available, with ordinary Step Functions and AgentCore Gateway charges.

The connector documentation lists ten supported Step Functions operations:

  • State-machine operations: StartExecution, StartSyncExecution, ListExecutions and DescribeStateMachine.
  • Execution operations: DescribeExecution, StopExecution and GetExecutionHistory.
  • Task-token operations: SendTaskSuccess, SendTaskFailure and SendTaskHeartbeat.

The first four operations can be pinned to one state machine in the target configuration. AWS says the pinned state-machine identifier is omitted from the agent-visible tool schema, and the gateway rejects an attempt to replace it. The remaining operations accept an execution ARN or task token supplied at call time and depend on IAM and possession of the relevant identifier.

This is a released connector, not a claim that every AWS account, region, agent framework or workflow is ready without configuration. Buyers still need to verify AgentCore Gateway availability, prerequisites, account policies, tool discovery, connector behavior, quotas and commercial terms in their selected region and account.

Separate documented behavior from business implications

What the AWS release supports—and what it does not prove
AreaAWS-documented behaviorPractical implicationNot established
Workflow sequenceThe agent chooses when to call and supplies input; the state machine controls its defined steps.Approved deterministic steps can remain outside the model's planning loop.That the trigger or input is correct, or every state-machine step is safe.
Existing workflowsStandard and Express state machines can be exposed without rewriting them as an MCP server.A team may reuse an existing AWS workflow.That a workflow designed for a trusted application is safe for agent-triggered use.
Tool schemaThe gateway creates tool schemas for selected operations and hides pinned parameters.The exposed interface can be narrower than a general AWS API client.That descriptions, input semantics or business constraints are sufficient for a model to call correctly.
IdentityThe connector calls Step Functions with the gateway service role.Access can be centralized and restricted with IAM.That the downstream workflow automatically knows or enforces the initiating human's identity and authority.
HistoryStandard workflows keep execution history; Express history depends on CloudWatch logging.A team can design evidence around the selected workflow type.That logs are complete, appropriately retained, free of sensitive data or sufficient for the business decision.
CostStandard Step Functions and AgentCore Gateway rates apply.The connector is consumption-based rather than a single flat feature.The complete cost of the model, gateway, workflow, logs, downstream services, engineering and review.

Write a workflow authority contract before exposing a tool

An agent-facing workflow should have a short, reviewable contract:

  • Purpose: one allowed business outcome and explicit non-goals.
  • Initiator: the people, agents, environments and gateway identities allowed to request it.
  • Workflow: exact account, region, state machine, alias or version and deployment owner.
  • Input: allowed fields, formats, classifications, size limits and trusted sources.
  • Authority: which records, customers, departments, services or dollar ranges are in scope.
  • Actions: reads, drafts, reversible writes and irreversible or external actions.
  • Approval: who must approve, what evidence they see and where the approval is enforced.
  • Limits: maximum executions, concurrency, duration, retries, records and spend.
  • Result: durable system of record, read-back test and evidence returned to the agent.
  • Stop and recovery: denial, invalid input, duplicate request, timeout, partial success, cancellation, compensation and manual fallback.

A state machine that accepts an unrestricted JSON object is not a bounded business tool merely because its steps run in a fixed order. The contract must be implemented in IAM, gateway policy, input validation, state-machine logic and downstream permissions—not only in a tool description or system prompt.

Use Orbital's AI automation readiness checklist to map the current task, data, approvals and exceptions before selecting the connector.

Scope each operation and the gateway service role

Do not enable all ten operations by default. Start with the minimum tool set required for one workflow.

Questions for each operation class
Operation classPrimary controlBuyer test
Start or inspect a state machinePin the exact state machine in the target and scope IAM to that state-machine ARN.Attempt to substitute another state machine and confirm the gateway rejects it.
Describe, stop or read an executionScope the gateway role to execution ARNs belonging to the approved state machine.Try an execution ARN from an unapproved state machine and confirm denial without leaked details.
Task-token callbacksProtect the token, expose only required callback operations and validate the returned result inside the workflow.Try an expired, reused, wrong-workflow and unauthorized token.
List operationsLimit what lists can reveal and whether listing is needed at all.Confirm the result cannot enumerate unrelated workflows or executions.

AWS documents that the gateway's service role authorizes Step Functions calls. State-machine operations can be restricted to named state-machine ARNs. Execution operations use a different resource type and can be scoped to executions of the selected state machine. Task-token callback actions do not support resource-level IAM permissions and therefore require broader Resource: * permission when enabled.

That last point changes the threat model. Do not infer that a gateway role is least privilege simply because its policy uses the narrowest resource syntax AWS supports. Review the combination of allowed actions, tool exposure, token delivery, gateway callers and workflow validation.

Treat Step Functions task tokens as secrets

A Standard workflow can pause on a callback task and issue a task token. An external worker or agent later returns success, failure or a heartbeat using that token. AWS says the three task-token operations cannot be restricted to specific executions through IAM; a caller also needs a valid token.

For any agent callback design:

  • deliver the token only to the component that must complete the task;
  • never place it in prompts, logs, analytics or error messages visible beyond that boundary;
  • bind the expected task, case, tenant and result schema inside the state machine;
  • reject a result that does not match the expected execution and business object;
  • set a timeout and document what happens when the agent never returns;
  • treat reuse, exposure or unexpected heartbeat activity as a security event;
  • rotate or disable the path by changing the workflow and gateway configuration when needed; and
  • avoid callbacks entirely when a simpler start-and-check pattern meets the business need.

Express workflows do not support the callback pattern described by AWS. Choose Standard or Express based on execution duration, history, callback and delivery requirements—not only latency.

Validate inputs and prevent duplicate business actions

The agent chooses the tool and supplies the non-pinned arguments. The state machine should assume that input can be incomplete, stale, duplicated, malicious or simply wrong.

  • Use a narrow schema with required fields, enumerated actions and explicit maximum lengths.
  • Resolve customer, order, case or project identifiers from an approved system rather than trusting free text.
  • Include the initiator, source, correlation ID, requested action and approval context as structured fields.
  • Reject unknown fields and out-of-scope tenants, services, amounts, destinations or dates.
  • Use idempotency keys or a durable request ledger before any action that must not repeat.
  • Check current record state immediately before a write; do not rely on the conversation's earlier snapshot.
  • Separate validation from execution and return a reviewable proposed action before an irreversible step.
  • Read the final state from the system of record and compare it with the intended result.

Step Functions provides retries and error handling, but a technical retry can repeat a downstream side effect unless the workflow and destination are designed for idempotency. Record which states are safe to retry, which require a unique business key and which require a compensation step.

Use the business software integration checklist for field maps, systems of record, reconciliation and duplicate handling.

Put material approvals inside the controlled workflow

Do not depend on the agent to remember when approval is required. Put the approval gate in the state machine or a downstream system that the agent cannot bypass.

An approval record should show:

  • the exact workflow, environment and business object;
  • the proposed fields, amount, destination and side effects;
  • the source evidence and any unresolved conflict;
  • the requesting identity and agent session or correlation ID;
  • the approver's identity, authority and decision time;
  • the approved scope and expiration;
  • the execution ARN and final system-of-record result; and
  • the rejection, timeout and escalation path.

Keep customer messages, payments, refunds, access changes, deletion, legal commitments and high-impact employment, financial, health or eligibility actions behind explicit human review unless the organization has established a narrower, independently enforced policy. A model's confidence score is not authorization.

Map data, logs and evidence before enabling production

A tool call can place business data in several surfaces: agent context, MCP request, gateway logs, Step Functions input and output, execution history, CloudWatch, downstream services and support systems. Map each copy, owner, purpose, retention and access role.

AWS says Standard workflows record execution history in Step Functions. Express workflows do not record execution history there; teams must configure CloudWatch Logs to view history and results. AWS also says Express log delivery is best effort, so a team needing guaranteed history should store the required business evidence in an appropriate durable system.

AWS documents that execution-history events may contain input, output and assigned variables. AgentCore Gateway can also log request and response bodies when log delivery is configured. Those capabilities are useful for debugging and can also duplicate sensitive payloads.

  • Send identifiers or references instead of full customer records where practical.
  • Exclude secrets, credentials, task tokens and unnecessary personal data from prompts and logs.
  • Configure log destinations, access, encryption, retention and deletion deliberately.
  • Separate operational telemetry from the durable business decision record.
  • Test whether truncation, sampling or best-effort delivery creates evidence gaps.
  • Enable and verify CloudTrail coverage needed for gateway and Step Functions API activity.
  • Define who investigates a mismatch between agent response, execution history and downstream state.

Do not claim auditability because a dashboard exists. Run a representative request from authenticated caller through final record and prove that the organization can reconstruct the trigger, input, approval, execution, side effects and accepted result.

Design for partial success, cancellation and recovery

Stopping an execution does not necessarily reverse work already completed in downstream systems. A workflow can fail after creating a ticket, sending an internal notification or reserving a resource. Build a failure register before launch:

  • invalid or adversarial input;
  • wrong customer, environment or state machine;
  • permission denial or changed IAM policy;
  • gateway timeout, throttling or unavailable target;
  • workflow timeout or execution-history limit;
  • partial downstream success;
  • duplicate start after an ambiguous response;
  • expired, leaked or reused task token;
  • approval timeout or conflicting approvers;
  • stale record between validation and write;
  • failure to deliver logs or evidence; and
  • manual takeover after the agent session ends.

For each failure, record prevention, detection, owner, stop condition, compensation, reconciliation and customer-impact rule. Test the manual path before production. A process is not resilient if the team cannot complete or unwind it without the agent.

Model complete operating cost

AWS's current AgentCore pricing page lists consumption-based Gateway charges for API invocations and separate charges for search requests and indexed targets used by semantic search. The Step Functions workflow, model, logs, data transfer and downstream AWS services have their own prices. Published rates and regions can change, so verify the account's price pages and contract at procurement time.

Estimate at least:

  • agent model input, output, retries and evaluation traffic;
  • gateway tool discovery, invocation, health checks and search;
  • Step Functions state transitions or Express duration and requests;
  • Lambda, queues, databases, APIs and other workflow services;
  • CloudWatch logs, metrics, traces, retention and queries;
  • network transfer and cross-region architecture;
  • security, identity, key management and secrets services;
  • engineering, testing, monitoring, incident response and support;
  • human approval and exception-handling time; and
  • migration, offboarding and manual fallback.

Measure cost per accepted business result, not per successful tool call. A low-cost execution that creates duplicate work or requires extensive review is not an efficiency gain. Use Orbital's AI automation ROI checklist to build a baseline and stop-or-scale decision.

Choose one bounded workflow for a 30-day pilot

A suitable first workflow is internal, reversible, low-volume and easy to verify. Examples include starting a data-quality check, collecting an approved operational report or creating a draft internal work item that a person reviews before external action.

Avoid beginning with customer messages, refunds, payments, access administration, deletion, production remediation or a workflow whose success cannot be read from a durable system of record.

  1. Days 1–7 — define. Select one workflow. Verify current regional availability and account prerequisites. Write the authority contract, baseline, data map, cost model, acceptance tests and stop conditions.
  2. Days 8–14 — restrict. Use a non-production environment and synthetic records. Expose only the required operation, pin the state machine, scope the gateway role, configure inbound authorization and protect identifiers and tokens.
  3. Days 15–21 — break it. Test wrong identity, substituted state machine, out-of-scope execution ARN, invalid input, duplicate start, stale state, approval rejection, timeout, partial failure, token misuse, log gap and manual recovery.
  4. Days 22–30 — shadow. Compare accepted results, reviewer time, exceptions, latency, full cost and recovery effort with the current process. Keep material production actions behind an independent approval gate.

Stop rather than expand if the agent can reach unrelated executions, the reviewer cannot see downstream effects, repeated requests duplicate work, logs expose sensitive payloads, evidence cannot connect the caller to the final result or the measured review burden exceeds the benefit.

Orbital's business take

The strongest part of this release is not MCP by itself. It is the ability to place a deterministic workflow behind a narrow agent tool without rewriting the workflow as a custom MCP server. That can preserve approved order, retries and long-running execution while giving an agent a simpler interface.

The risk moves to the boundary: who can call, which operation is exposed, what input the agent supplies, which identity reaches Step Functions, what the workflow can change and whether the organization can prove the final result. “The agent cannot reorder the states” is valuable, but it does not answer those questions.

Explore Orbital's AI automation services, tech stack consulting, connected-business planning, automation and technology strategy, owner oversight systems and service-business planning. A 60-minute consultation can turn one workflow into an evaluation brief; AWS architecture, implementation, model and service usage, security, privacy, legal review, monitoring and ongoing support are separately scoped.

Sources and editorial limits

Primary sources were checked November 3, 2026. AWS dated the connector release October 29 and says it is available in regions where AgentCore Gateway is available. AWS documentation supplies the connector operations, pinned-parameter behavior, gateway service-role model, IAM limitations, authentication options, observability behavior, Step Functions history differences and AgentCore pricing structure. These are vendor statements. The authority contract, control matrix, task-token checklist, input and replay controls, approval record, failure register, cost model and 30-day pilot are Orbital editorial recommendations—not AWS requirements. Orbital did not configure or test the connector, access an AWS account, invoke a workflow, audit IAM or logs, reproduce vendor claims, measure latency or cost, review account-specific pricing or contracts, or determine a particular organization's legal, security, privacy or compliance requirements. Public sources do not establish universal setup effort, complete operating cost, customer-specific quotas, performance, service levels or business outcomes. This article does not guarantee availability, savings, security, compliance or results.

Primary source links

Your next chapter.
Let’s give it lift.

Start a conversation