Setting the file. One moment.
Subchapter 25.67
references/vendored/ai/ai-migration-guardrails.mdMarkdown10 KBView on GitHub
Canonical guardrails for every agentic migration path. Source-cloud agnostic: these are Bedrock/AgentCore constraints, not source-provider facts. Vendored into each consuming skill as
references/vendored/ai/ai-migration-guardrails.mdand kept byte-identical byshared:check; edit HERE, then runshared:sync.
Shared warnings and constraints for all agentic migration paths. Loaded once by design-ai.md when agentic_profile.is_agentic == true. Path-specific design references (Harness, Strands, retarget) should NOT duplicate these — reference this file instead.
AgentCore services have different regional footprints. Always validate via aws___get_regional_availability from the AWS MCP Server before recommending.
As of July 2026:
| Service | Availability | Regions |
|---|---|---|
| AgentCore Runtime (GA) | All commercial regions | us-east-1, us-east-2, us-west-2, us-west-1, ap-southeast-1, ap-southeast-2, ap-northeast-1, ap-northeast-2, ap-south-1, eu-central-1, eu-west-1, eu-west-2, eu-north-1, sa-east-1, ca-central-1 |
| AgentCore Harness (GA) | All commercial regions | Same as Runtime (GA June 2026) |
| AgentCore Memory (GA) | All commercial regions | Same as Runtime |
| AgentCore Gateway (GA) | All commercial regions | Same as Runtime |
| AgentCore Policy (GA) | 13 regions | us-east-1, us-east-2, us-west-2, us-west-1, ap-southeast-1, ap-southeast-2, ap-northeast-1, ap-south-1, eu-central-1, eu-west-1, eu-west-2, eu-north-1, ca-central-1 |
IMPORTANT: These lists go stale. The aws___get_regional_availability call (AWS MCP Server) is the source of truth. Use the table above only as a fallback if the call fails.
If target region is unavailable for a recommended service:
aws-design-ai.json → regional_warnings[]Bedrock Mantle serves OpenAI-compatible and Anthropic-compatible APIs on Bedrock.
Inference on bedrock-mantle for these models is governed by two per-model, per-region quotas: input tokens per minute and output tokens per minute. There is no requests-per-minute quota. Exceeding a TPM quota returns HTTP 429. Cached input tokens read through prompt caching do not count against the input-TPM quota.
| Workload Volume | Risk Level | Guidance |
|---|---|---|
| Low | Low | Default TPM quotas are ample |
| Medium | Medium | Monitor 429s against token throughput, not request rate; enable prompt caching |
| High | High | Enable prompt caching first (cached input is exempt from input TPM), then request a quota increase |
The bedrock-runtime fallback exists only for GPT-5.6. GPT-5.5 and GPT-5.4 are bedrock-mantle only and in-region only — for them, “switch to bedrock-runtime“ requires moving to a different model (Bedrock-native or gpt-oss), a model change with its own eval cost. GPT-5.6 Sol/Terra/Luna DO have a bedrock-runtime path via CRIS inference profiles (us./in./global. prefixed ids; the model cards recommend runtime for new applications) — a legitimate endpoint option with its own quota family, and on Global CRIS it is also the cost-parity option. Scaling levers on the mantle path, in order:
max_retries on the OpenAI SDK)See references/shared/openai-on-bedrock.md for the endpoint, region matrix, and caching parameters.
For non-OpenAI models served through Mantle, verify current quota dimensions and any shared-account limits in the Mantle documentation (opens in a new tab) before advising on throughput. Do not carry the OpenAI TPM-only model over to other providers without checking, and do not assume previously documented shared-RPM behavior still applies.
Claude models on Mantle have an additional output TPM cap that differs by model generation:
| Model Family | Output TPM Cap | Notes |
|---|---|---|
| Claude 4.7+ | 2,000,000 output TPM | Per-model cap applies |
| All other Claude models | No per-model output TPM limit | Standard account TPM limits apply |
Impact for migration decisions:
bedrock-runtime/CRIS path is an available alternative (its own quota family), while GPT-5.5/5.4 have no runtime pathgpt-oss targets: these do run on bedrock-runtime, so standard account TPM limits and the Converse-path mitigations applybedrock-runtime for production only when the target model actually has a bedrock-runtime pathagentic_profile.tools contains write-capable tools (database writes, API mutations, file operations) or when the startup has compliance requirements.Before recommending any Bedrock model in an agentic design:
references/vendored/ai/ai-model-lifecycle.md for model statusFor agentic workload cost estimation:
references/shared/pricing-cache.md (±5-10% accuracy)references/shared/pricing-fallback.md (cache-miss → estimated/unavailable)AgentCore Runtime and Harness pricing: consumption-based, no upfront cost. Include in estimate only if the user selects Harness or Strands path.
Do NOT output fixed week estimates for agentic migrations. Output ranges with drivers:
Format: “[low]–[high] weeks depending on [driver 1] ([value]), [driver 2] ([value]), [driver 3] ([value])”
Drivers to include:
agentic_profile.agent_count)agentic_profile.tool_count)agentic_profile.orchestration_pattern)Example: “2–5 weeks depending on agent count (3), tool count (8), and graph complexity (hierarchical with conditional routing)”