AWS AppConfig AI Configurations: Power-User Guide to Models and Prompts

Last updated: November 20, 2026

AWS AppConfig AI configurations, launched November 19, 2026, add two profile types: AI model invocation configurations for model IDs, inference parameters and messages, and AI prompt configurations for versioned prompts. You can change both without redeploying code, roll changes out gradually, and roll back automatically on alarms, at no additional cost.

Contents

What changed on November 19

In its What’s New post dated November 19, 2026, AWS announced two new configuration profile types for generative AI applications in AWS AppConfig. Teams can change model settings and prompts while an application runs, without a code deployment.

Both types use the deployment, retrieval and rollback features AppConfig already offers for feature flags and freeform configurations.

If you are building agents on AWS, this sits alongside other recent launches we have covered, such as Amazon OpenSearch Memory for agent memory.

The two new profile types at a glance

Item AI model invocation configuration AI prompt configuration
Type name AWS.AppConfig.AIModelInvocation AWS.AppConfig.AIPrompt
What it stores Model identifier, inference parameters, messages and custom parameters A single prompt as plain text
Format JSON (application/json) Any text format, including Markdown (text/plain)
Validation Checked against the AI model invocation schema at creation and deployment Stored as text
Where it lives AppConfig hosted configuration store AppConfig hosted configuration store
Best for Changing models, tuning parameters, few-shot examples Iterating on prompts on their own schedule and reusing them across applications

The two types work together. According to the AI prompt configuration guide, prompts tend to change far more often than model selection or inference parameters, so keeping them separate lets you update a prompt without redeploying the model configuration.

Inside the AI model invocation schema

The AI model invocation configuration guide describes a strongly typed schema for fields that are part of the model call, common across AI providers and possible to validate.

  • schemaVersion (required): the schema version, so the schema can evolve without a new profile type.
  • model (required): the model identifier, for example anthropic.claude-sonnet-5.
  • inferenceParameters: temperature (minimum 0), topP (0 to 1), maxTokens (minimum 1) and stopSequences. These mirror the InferenceConfiguration used by the Amazon Bedrock Converse API.
  • messages: messages you write to prime the model, such as a system prompt and user/assistant few-shot examples. The user’s real-time input is not stored.
  • customParameters: free-form JSON for provider-specific options such as extended thinking, reasoning effort or output format.

The AWS example sets a Claude model on Amazon Bedrock with a temperature of 0.7, topP of 0.9, maxTokens of 4000 and a 4000-token extended thinking budget in customParameters.

If you are watching model spending, our guide on how much the Claude API costs shows why a change to maxTokens or a thinking budget can move your bill.

How to create an AI model invocation profile

First create an AppConfig application to hold the profile. Then, in the console:

  1. Open the AWS Management Console, search for AWS AppConfig and choose Create.
  2. Choose AI model invocation configuration profile.
  3. Under Save under application resource, pick your application, then enter a configuration profile name and an optional description.
  4. In the AI model invocation definition section, enter the Model identifier.
  5. Optionally set temperature, top P, maximum tokens and stop sequences, add system or example messages, and enter custom parameters as JSON.
  6. Choose Next, review the settings, and choose Save and continue to deploy.
  7. Deploy the configuration to an environment so your application can read it.

From the AWS CLI, run aws appconfig create-configuration-profile with --location-uri hosted and --type AWS.AppConfig.AIModelInvocation, then upload your JSON with aws appconfig create-hosted-configuration-version.

How to create an AI prompt profile

  1. In the AppConfig console, choose Create, then AI prompt configuration profile.
  2. Choose the application, enter a profile name, and optionally add a description.
  3. Under Additional options, optionally choose a customer managed AWS KMS key and add tags.
  4. Choose Next and enter the prompt in the prompt editor, in Markdown or plain text.
  5. Review the details, save, and deploy.

On the command line, use profile type AWS.AppConfig.AIPrompt and content type text/plain.

How your application reads the settings

AppConfig stores and delivers the settings; it does not send the request to the model. Your application fetches the deployed configuration from AWS AppConfig Agent at a local HTTP endpoint, such as http://localhost:2772/applications/myapp/environments/prod/configurations/support-agent.

The model configuration comes back as JSON and the prompt as plain text. In the AWS examples, the application maps the system messages and model fields onto an Amazon Bedrock Converse request, passing customParameters as additional model request fields.

Gradual rollouts and automatic rollback

AWS gives the example of rolling a change to 10%, then 20%, then 30% of traffic while you measure operational metrics, customer experience and token usage.

AppConfig’s predefined deployment strategies include AppConfig.Linear20PercentEvery6Minutes, which reaches 20% of targets every six minutes over 30 minutes and then watches CloudWatch alarms for 30 more minutes, and AppConfig.Canary10Percent20Minutes, which grows by 10% steps over 20 minutes with a 10-minute watch period. AWS recommends both for production.

If an alarm fires during the rollout or the watch period, AppConfig rolls back to the previous version. With a separate prompt profile, that means the previous prompt version, not your model settings.

Pricing and availability

AWS says both profile types are available in all AWS Regions where AppConfig is available, at no additional cost. The usual AppConfig pricing still applies to retrieval:

  • $0.0000002 per configuration request made through API calls
  • $0.0008 per configuration received

The AWS pricing example, with 2,000 targets checking every two minutes and one configuration updated three times a day, comes to $152.64 a month.

Limitations to plan for

  • No provider-side checks: AppConfig confirms customParameters is valid JSON but does not check what is inside, so a wrong provider option can pass validation.
  • Rollback needs alarms: automatic rollback depends on the alarms you configure.

For other AWS building blocks for agents, see our guides to the AgentCore Web Search API and AgentCore Web Answer for cited web answers.

Frequently asked questions

What are AWS AppConfig AI configurations?

They are two configuration profile types AWS added on November 19, 2026: AI model invocation configurations, which hold the model identifier, inference parameters and messages, and AI prompt configurations, which hold a prompt as its own versioned artifact. Both can change while your application runs.

Do AppConfig AI configurations cost extra?

No. AWS says both profile types are available at no additional cost. Standard AWS AppConfig pricing still applies when targets request and receive configuration data: $0.0000002 per configuration request and $0.0008 per configuration received.

Which AI models can I use with AppConfig?

The model field is an open string, so you enter the identifier your provider uses, such as anthropic.claude-sonnet-5 in the AWS example. Provider-specific options go in customParameters, which AppConfig checks only for well-formed JSON.

Does AppConfig call the AI model for me?

No. Your application retrieves the deployed configuration from AWS AppConfig Agent and maps the fields to the request format your AI provider expects, for example an Amazon Bedrock Converse call.

Can AppConfig roll back a bad prompt automatically?

Yes. If a configured alarm in Amazon CloudWatch, Datadog, New Relic or another observability tool goes off during a rollout, AppConfig can roll back to the previous version.

Sources

Image: AWS logo, public domain, via Wikimedia Commons. Resized. AWS is a trademark of Amazon.com, Inc. or its affiliates; FreeAnswers is not affiliated with Amazon.

Related answers

Leave a Comment