Step Functions as Agent Tools: A UK SRE Playbook for AgentCore Gateway and Agent Registry in London
AWS made Step Functions workflows usable as AgentCore Gateway tools on 29 October 2026, a day after Agent Registry arrived in London. Here is how I would let agents choose while workflows execute, with UK controls around it.

AWS made Step Functions workflows usable as AgentCore Gateway tools on 29 October 2026, a day after Agent Registry arrived in London. Here is how I would let agents choose while workflows execute, with UK controls around it.
- Step Functions Standard and Express workflows can be used as AgentCore Gateway tools from 29 October 2026.
- Agents choose when to run a workflow and with what input; Step Functions runs each step as defined.
- AWS Agent Registry became available in Europe (London) eu-west-2 on 28 October 2026.
- Expose one workflow per business action and validate agent input in the first state.
- Scope gateway and state machine IAM roles to named workflows only.
- Use task-token approvals where money, customer data or production impact is involved.
- Register approved tools in London with owner, data classification and runbook.
- Alarm on failed runs and budget per workflow, since Standard workflows bill per state transition.
What changed
Two AgentCore announcements landed this week that fit together neatly:
- 29 October 2026: AWS Step Functions workflows can now be used as tools for AI agents through Amazon Bedrock AgentCore Gateway. The agent decides when to run a workflow and what input to pass, and Step Functions runs each step exactly as defined. Existing Standard and Express workflows can be used, agents can look up workflow runs, and it is available in every Region where AgentCore Gateway is available at standard Step Functions and Gateway rates.
- 28 October 2026: AWS Agent Registry is now available in Europe (London) eu-west-2, along with Frankfurt, Ohio and Canada (Central). The registry is a managed catalogue for MCP servers, tools, agents and agent skills, with an approval workflow and a native MCP endpoint.
Put simply, you can now keep the risky, multi-step part of an agent's job in a deterministic workflow, and publish that workflow as an approved tool from a registry in London.
Why this matters for SREs
The hardest part of putting an agent near production is not the model. It is making sure the agent cannot improvise a process that has rules. AWS's own example is a refund: the agent starts the refund workflow, and Step Functions runs eligibility, manager approval and credit in the same order every time.
That is the pattern I would use for anything with a policy behind it: the agent chooses, the workflow executes. Retries, timeouts, compensation and human approval live in the state machine, where they are versioned, tested and visible in execution history, instead of in a prompt.
A UK playbook
1. Choose the right workflows to expose
Good first candidates are processes that already exist and already have clear inputs: refunds, account unlocks, environment start-up and tear-down, evidence collection for an incident. Avoid exposing generic "run anything" workflows. One workflow per business action keeps the tool description honest and the audit trail readable.
2. Validate input at the front of the workflow
The agent chooses the input, so treat it as untrusted. Make the first state a validation step with a strict schema, reject unexpected fields, and fail closed. Never let free text flow straight into a downstream API call.
3. Pick Standard or Express deliberately
Standard workflows suit long-running, auditable processes with human approval steps. Express workflows suit short, high-volume tasks. If a person or a regulator might later ask what happened on a given run, prefer Standard for its execution history.
4. Scope the gateway and the role
Give the gateway target an IAM role that can start only the named state machines, and give each state machine a role that can touch only what its steps need. Read-only run lookups for an operations agent should be a separate target from anything that starts a workflow.
5. Keep humans in the loop where policy says so
Use a task-token approval step for actions with money, customer data or production impact. The agent can request the action; a person still approves it. That keeps separation of duties intact, which auditors in UK financial services will ask about.
6. Register tools in London
With Agent Registry now in eu-west-2, UK teams can keep their catalogue of approved MCP servers, tools and agents in-region. Use the approval workflow so a workflow becomes discoverable only after review, and record the owner, data classification and runbook for each entry. Confirm in the AWS documentation which data the registry stores before relying on it for residency claims.
7. Make runs observable
Agents can now look up workflow runs, which is useful for on-call. Pair that with alarms on failed executions and a dashboard that separates agent-started runs from human-started ones, so you can spot an agent looping or calling a workflow far more often than expected.
8. Watch the bill
You pay standard rates for Step Functions and AgentCore Gateway. Standard workflows are charged per state transition, so a chatty agent calling a long workflow repeatedly can add up. Set budgets and alarms per workflow before go-live.
Bottom line
This is one of the more useful agent launches of the year for production teams, because it moves control back into tools SREs already trust. Let the agent decide what needs doing, let Step Functions decide how, register the result in London, and keep a human approval where your policy needs one.
This launch moves control back to where SREs already have it. The agent decides what needs doing; the state machine decides how, with retries, approvals and history you can test and audit. That is a far safer shape than encoding process rules in a prompt.
