canonical: https://jentic.com/apis/amazonaws.com/amazon-augmented-ai-runtime

# AWS Amazon Augmented AI Runtime

Jentic publishes the only available OpenAPI specification for Amazon Augmented AI Runtime, keeping it validated and agent-ready. Amazon Augmented AI (A2I) is the human-in-the-loop layer for ML predictions - when a model's confidence falls below threshold, A2I routes the prediction to a configured workforce (private, vendor, or Amazon Mechanical Turk) for human review. The runtime API starts, monitors, stops, and deletes individual human review tasks (called HumanLoops) tied to a flow definition you create in the SageMaker console.

## For AI agents

Send low-confidence ML predictions to human reviewers and retrieve the human-validated result, so agents can defer hard cases instead of guessing.

## Scope

Does not create flow definitions, manage workforces, or run inference itself - use for starting, reading, stopping, and deleting individual human review tasks against an existing flow definition only.

## Capabilities

- Start a human review task with StartHumanLoop attached to a flow definition and input data
- Read the status and reviewer output of a human review task with DescribeHumanLoop
- List human review tasks created against a given flow definition with ListHumanLoops
- Stop an in-progress human review task with StopHumanLoop when the result is no longer needed
- Delete a completed human review task with DeleteHumanLoop to clean up state
- Pass arbitrary structured input to a human review task and receive structured reviewer output

## Use cases

### Human-in-the-loop for low-confidence predictions

ML teams use A2I to defer borderline predictions to human reviewers instead of acting on them blindly. When a SageMaker endpoint returns a confidence score below threshold, the application calls StartHumanLoop with the input data and the FlowDefinitionArn that defines the reviewer UI and workforce; once a reviewer submits, DescribeHumanLoop returns the structured human-validated result.

Example prompt: Call StartHumanLoop with FlowDefinitionArn for the document-classification flow, HumanLoopName='doc-1234', and the document JSON as input, then poll DescribeHumanLoop until status is Completed.

### Compliance and audit review of ML decisions

Regulated teams use A2I to ensure a percentage of automated decisions are reviewed by qualified humans for audit. By configuring the flow definition to sample a fixed proportion of predictions and route them to a private workforce, A2I produces a defensible record of human-validated outcomes that ListHumanLoops and DescribeHumanLoop can replay during audits.

Example prompt: Call ListHumanLoops with FlowDefinitionArn and a CreationTimeAfter filter to retrieve all sampled review tasks from the last quarter for audit export.

### Document and image review for ML pipelines

Document- and image-processing teams use A2I to add human verification to OCR, object-detection, and form-extraction pipelines. The flow definition renders the document or image alongside the model's prediction in the reviewer UI; the reviewer corrects or confirms it; DescribeHumanLoop returns the corrected result for downstream pipeline steps.

Example prompt: Submit StartHumanLoop with the Textract output as InputContent and the document-extraction flow definition, then on Completed status feed reviewer-corrected fields back into the downstream pipeline.

### Agent-driven human-fallback through Jentic

AI agents that need to defer hard cases use Jentic to call StartHumanLoop and DescribeHumanLoop without holding raw AWS credentials. Jentic stores keys in your Jentic One instance, signs each request with SigV4, and returns the structured response - so an agent can pause its workflow until a human verdict arrives, then resume with the validated result.

Example prompt: Search Jentic for 'send a prediction for human review', load StartHumanLoop, and execute it when the agent's own confidence is below 0.7, then poll DescribeHumanLoop until Completed.

## Key endpoints

| Method | Path | Description |
| --- | --- | --- |
| POST | /human-loops | Start a human review task |
| GET | /human-loops/{HumanLoopName} | Read the status and output of a human review task |
| GET | /human-loops | List human review tasks for a flow definition |
| POST | /human-loops/stop | Stop an in-progress human review task |
| DELETE | /human-loops/{HumanLoopName} | Delete a completed human review task |

## Key resources

- **HumanLoop** — An individual human-review task identified by HumanLoopName
- **FlowDefinition** — Defines the reviewer UI, workforce, and conditions - created out-of-band in SageMaker
- **InputContent** — The structured payload sent to the reviewer (model prediction, document, image)
- **HumanLoopOutput** — The reviewer's structured response written to S3 on completion

## Why Jentic

- **Setup:** Wiring Amazon Augmented AI Runtime by hand means implementing AWS SigV4 signing, resolving the a2i-runtime.sagemaker regional host, and reading reviewer output from the S3 URI the loop writes rather than an inline response. Through Jentic you install once, import Amazon Augmented AI Runtime from the API Directory, store your AWS access key and secret once, and your agent calls it.
- **Permission scoping:** A2I Runtime takes the FlowDefinitionArn and input payload in the request body, so scope your agent to the operations it needs rather than to a single flow: for example allow StartHumanLoop and DescribeHumanLoop for reading verdicts. You choose the operations it may call, so StopHumanLoop and DeleteHumanLoop are not included unless you add them.
- **Credential handling:** Your AWS access key and secret are stored once, encrypted, by your own Jentic One instance and each request is signed with SigV4 at execution time. They never enter the agent's prompt, logs, or context.
- **Discovery method:** Agents search Jentic by intent such as 'send a prediction for human review' or 'check whether a human loop completed', and Jentic returns the matching operation like StartHumanLoop or DescribeHumanLoop with its input schema so the agent calls it without browsing the AWS docs.

## Related APIs

- **Amazon SageMaker** — Create the FlowDefinition and produce the predictions that A2I reviews
- **Amazon Textract** — Textract output is a common A2I input for human document review
- **Amazon Mechanical Turk** — Lower-level crowdsourcing API - A2I sits on top of it for ML-specific review

## FAQ

### Why is there no official OpenAPI spec for Amazon Augmented AI Runtime?

AWS does not publish an OpenAPI specification. Jentic generates and maintains this spec so that AI agents and developers can call Amazon Augmented AI Runtime via structured tooling. It is validated against the live API and kept up to date. Get started with Jentic One, the self-hosted execution layer.

### What authentication does the Amazon Augmented AI Runtime API use?

A2I Runtime uses AWS Signature Version 4 HMAC signing with an access key ID and secret access key, optionally with a session token. Through Jentic, those credentials live in your Jentic One instance and the agent never handles raw secrets - Jentic signs each request server-side.

### Can I create a flow definition with the Amazon Augmented AI Runtime API?

No - flow definitions are created via the SageMaker control-plane API (CreateFlowDefinition), not the A2I runtime. The runtime only starts, reads, stops, and deletes individual HumanLoop tasks against an already-defined flow. Set up the flow definition first in SageMaker, then call StartHumanLoop here.

### What are the rate limits for the Amazon Augmented AI Runtime API?

AWS does not publish a fixed RPS for A2I Runtime in the spec; standard AWS API throttling applies. The deeper bottleneck is reviewer throughput, not API rate - high-volume workflows should size the workforce appropriately and use ListHumanLoops to monitor backlog.

### How do I send a model prediction for human review through Jentic?

Search Jentic for 'send a prediction for human review', load StartHumanLoop, and call POST /human-loops with FlowDefinitionArn, a unique HumanLoopName, and the input JSON. Then poll DescribeHumanLoop until HumanLoopStatus is Completed and read HumanLoopOutput.OutputS3Uri for the reviewer's verdict. Install with pip install jentic.

### Where does the Amazon Augmented AI Runtime API store the reviewer's output?

Reviewer output is written to the S3 bucket configured in the flow definition, not returned inline. DescribeHumanLoop returns an OutputS3Uri pointing at the JSON object the reviewer produced; your application reads it directly from S3 once the loop status is Completed.

### Can I limit what my agent is allowed to do with the Amazon Augmented AI Runtime API?

Yes. Because you run Jentic One yourself, your own rules decide which A2I Runtime operations the agent may call, so you can allow StartHumanLoop and DescribeHumanLoop for sending predictions and reading verdicts while leaving StopHumanLoop, DeleteHumanLoop, and ListHumanLoops out unless you add them. The FlowDefinitionArn and input payload travel in the request body, so scoping is done at the operation level rather than per flow. Your AWS access key and secret stay in your own Jentic One instance and are signed with SigV4 at execution time, so the agent only ever exercises the operations you granted.