Article

Choose Skyflow data controls for AgentCore workflows

For teams building customer-data assistants, choose which Skyflow control fits each recipient and check preview limits, setup and separate costs.

Editorial illustration for Choose Skyflow data controls for AgentCore workflows: a controlled task flows from input to output. Not documentary evidence.

AWS’s November 19, 2026 Skyflow–AgentCore walkthrough gives teams building customer-data assistants three places to transform sensitive information. Shield a tool when it does not need customer identities; route model calls through the gateway when the model should not receive them; use the document preview when a reader needs a de-identified copy.

The key decision is which recipient needs the original data: a tool, model or document reader. The agent application—the software coordinating those calls—is a separate recipient. Restoring a customer’s name to that application does not mean the name stays out of its next model call. The reference architecture makes those paths distinct.

When should you hide identities from a tool?

Choose tool shielding when the external tool can work with substitutes for selected customer details. The reference implementation replaces selected input values with reversible vault tokens: substitutes backed by originals stored in Skyflow. Tokens in returning results can then be restored before the application reads them.

That return step can be useful when staff need a readable answer containing the customer’s details. But it also puts original values back into the application. If the application sends that answer to a model, protecting the earlier tool call alone is insufficient.

There is a functional tradeoff, too. A tool that needs an actual customer identifier to look up a record may not work with a substitute. Decide what the tool genuinely needs before choosing which fields to transform; the sample policy supports field selection and per-tool rules.

What must change to protect a model call?

Model shielding requires its own route. Add a Gateway inference target and point the application’s model client at that endpoint. The Skyflow inference policy then determines which prompts, conversation turns, tool arguments, tool results and supported attachments are transformed before reaching the provider.

In the supplied configuration, the returning answer restores only tokens created for that request. This allows an application to receive a readable result while the selected details sent to the model were substitutes. It does not remove originals already held elsewhere in the application.

Check compatibility before choosing this path. The reference documentation’s limitations include supported request formats, buffered rather than word-by-word shielded answers, and tools run by the model provider that sit outside this transformation path. A provider’s own search or retrieval result is not automatically covered.

Keeping information out of a prompt is another distinct choice. Our guide to keeping large tool results out of AgentCore prompts addresses selective retrieval and context size, not de-identification of the values you do send.

Can you restore a redacted document?

First check access. The AWS walkthrough describes string de-identification and restoration as generally available, but signed-URL document processing uses Skyflow Detect v2 in preview. The reference documentation says that capability must be enabled for the vault; a general Skyflow subscription does not establish access.

For a file URL matched by the policy, Skyflow fetches the original and the interceptor replaces its link with one pointing to a transformed copy. The original bytes do not pass through the interceptor’s Lambda function, but Skyflow still processes them. This is a link substitution, not deletion of the source document.

  • Supported text and Office files: TXT, JSON, CSV, XML, JSONL, DOCX, XLSX and PPTX can retain reversible vault tokens. Authorized recovery remains possible outside this file-handling path; the gateway does not automatically restore the file on its return journey.

  • PDFs and images: the output uses one-way substitutions. You cannot recover the replaced values from that transformed copy. Keep this distinction in mind if a later authorized reader will need the originals.

Those format and restoration rules make the output format part of the decision, not merely a presentation preference.

The shipped document policy allows one file, sets a 512 KiB size threshold and names particular response fields and permitted storage URLs. These are sample settings, not universal service limits. Adapt them to the actual tool response; unmatched file references are not automatically transformed.

The size check in the handler rejects a known oversized file, but processing can proceed when its size is unknown. Avoid treating the sample as either an unlimited-file service or a guarantee that every oversized file is rejected before processing.

What does blocking on failure actually cover?

The sample defaults to fail closed: if a transformation fails, the affected call or result is blocked. Its configurable fail-open alternative can pass the original through. These failure settings determine what happens when processing fails—not whether detection found every sensitive value.

Coverage still depends on the selected entity types, fields, length thresholds and exceptions in the policy. Review logs separately, too: the repository’s logging guidance warns that enabled Gateway application logging can record payloads. Configuring the interceptor’s logs does not configure every application log, trace or memory store.

Check existing integrations before installation. AWS permits at most one request interceptor and one response interceptor. If either slot already handles authorization or auditing, design how that behavior will be combined with the transformer; a second same-direction interceptor cannot simply be stacked alongside it.

Content transformation also does not replace AgentCore private API routing or per-user versus service access to Snowflake. How traffic travels, who may retrieve a record and which recipient sees its original values are separate decisions.

What access, regions and costs must you confirm?

The integration code has an MIT license, but that is not a free Skyflow service plan. Deployment requires AWS configuration and a Skyflow vault with the appropriate Detect capability. Ask for confirmation of preview access, processing and storage locations, capacity and token-retention arrangements.

As of November 19, AWS lists Gateway in both Mumbai and Hyderabad in its regional availability matrix. That does not establish availability of every other AgentCore feature, your chosen model or Skyflow processing in the same location. Check the complete workflow’s regional requirements.

Skyflow’s AWS Marketplace listing uses private offers with custom quotes. An exact service price, included detection volume and document-preview entitlement require confirmation; none can be inferred from the open-source code.

Illustrative Gateway subtotal: AWS’s published pricing lists ordinary Gateway API invocations at $0.005 per 1,000. Assuming one million such invocations in a month, the arithmetic is 1,000,000 ÷ 1,000 × $0.005 = $5 for that line alone.

That is not a monthly price for the integrated assistant. Budget separately for Skyflow, model inference, Lambda execution, state storage, logs, document storage and transfer, plus Gateway search and indexing if used. Confirm applicable regional rates and contract terms.

Start with one workflow and identify who needs originals at each step. Then use approved synthetic examples in a pilot to check recipient-visible data, missed detections, task usefulness, failure handling and latency. Choose tool, model and document controls from those requirements, and price the complete workflow rather than a single low-cost API line.