What users are trying to accomplish
Build generative-AI platforms on AWS through Planton — the fastest-growing 2026 workload. Today the catalog has zero coverage of Amazon Bedrock, so an agent composing an AI architecture on AWS hits a hard stop and has to drop to hand-written Terraform.
Current state
No Bedrock kind of any kind exists. The Terraform AWS provider carries three Bedrock service packages that together are the surface to model:
- bedrock — foundation-model access, custom models, provisioned throughput, model invocation logging, guardrails, inference profiles.
- bedrockagent — agents, agent aliases, knowledge bases, data sources, prompts, flows.
- bedrockagentcore — the 2026 AgentCore platform: agent runtimes, gateways, memory, browsers, code interpreters, workload identities.
Scope
Model the Bedrock surface as a coherent family of first-class kinds (grouped sensibly — e.g. a knowledge-base kind that composes its data sources, a guardrail kind, an agent kind, AgentCore runtime/gateway/memory kinds), at full configuration parity with the provider schema, both IaC engines, chart-composable via StringValueOrRef references into the supporting storage / OpenSearch / IAM kinds. An AI-platform infra chart with Bedrock as the centerpiece is the natural flagship deliverable once the kinds are proven.
The scoping session must census all three provider packages before fixing the kind list — the block is larger than a single Bedrock package suggests.
What users are trying to accomplish
Build generative-AI platforms on AWS through Planton — the fastest-growing 2026 workload. Today the catalog has zero coverage of Amazon Bedrock, so an agent composing an AI architecture on AWS hits a hard stop and has to drop to hand-written Terraform.
Current state
No Bedrock kind of any kind exists. The Terraform AWS provider carries three Bedrock service packages that together are the surface to model:
Scope
Model the Bedrock surface as a coherent family of first-class kinds (grouped sensibly — e.g. a knowledge-base kind that composes its data sources, a guardrail kind, an agent kind, AgentCore runtime/gateway/memory kinds), at full configuration parity with the provider schema, both IaC engines, chart-composable via StringValueOrRef references into the supporting storage / OpenSearch / IAM kinds. An AI-platform infra chart with Bedrock as the centerpiece is the natural flagship deliverable once the kinds are proven.
The scoping session must census all three provider packages before fixing the kind list — the block is larger than a single Bedrock package suggests.