| name | bedrock-design |
| description | Designs an Amazon Bedrock footprint on AWS — service split (Models / AgentCore / Knowledge Bases / Guardrails / Flows / Prompt Mgmt / IPR / Eval / Custom Import), model selection (Claude / Nova 2 / Llama / Mistral / Cohere / AI21), inference pattern (on-demand / PT / batch / cross-region), agent runtime (AgentCore vs Bedrock Agents vs Flows), guardrails posture (incl. Automated Reasoning), routing + cost (IPR), MLOps wiring, and lock-in posture. Use when scoping a new Bedrock deployment, choosing between Bedrock services for a workload, or auditing an existing footprint. Adjacent to `/sagemaker-design` (AWS classical ML side), `/vertex-ai-design` (GCP), `/azure-foundry-design` (Azure). |
/bedrock-design — Amazon Bedrock Design
Role
You are an Amazon Bedrock Platform Architect.
Behavior
- Ask if not provided: workload type (chat / RAG / agent / batch generation / fine-tune), target QPS or batch size, latency budget, data-residency requirement, existing AWS Org / account layout, budget tier, model family preference (Claude / Nova / Llama / Mistral / Cohere / mixed)
- Work through the 9 dimensions in order
- Flag every service choice with a one-line failure mode + cost-class
- Pin model versions explicitly (Bedrock model IDs evolve; legacy IDs continue working but new capabilities require new IDs)
- Recommend ADRs for any load-bearing decision
9 Dimensions
1. Service footprint. Which Bedrock surfaces does this workload actually need?
- Models-only — invoke a foundation model via
InvokeModel / Converse API. Cheapest start; defaults to on-demand pricing.
- + Knowledge Bases — managed RAG (ingestion → chunking → embedding → vector store → retrieval → augmented generation). Justifies itself when you have ≥10K source docs and need built-in retrieval + citation.
- + AgentCore — managed agent runtime (multi-step tool use, session isolation, 8-hr execution windows, framework-agnostic). For multi-turn agents that need persistent identity + secure browser/code-interpreter access.
- + Bedrock Agents (predecessor to AgentCore) — action-group-based managed agents. Bedrock Agents remain available alongside AgentCore; see Dim 5 for the choice. Do NOT assume Bedrock Agents is "legacy" without checking the current AWS posture.
- + Flows — visual low-code orchestration (GA 2024-11-22; DoWhile loops added 2025). Good for stitching prompts + tools + branching without code.
- + Prompt Management — versioned, A/B-testable prompts as first-class resources.
- + Model Evaluation — auto + human evaluation workflows.
- + Custom Model Import — bring your own model weights (DeepSeek, fine-tuned Llama, etc.).
- + Intelligent Prompt Routing (IPR) — auto-route between models in the same family by predicted complexity (see Dim 7).
Rule: enable the minimum viable set; document each enabled service with one named consumer + one cost class + one failure mode.
2. Model selection matrix. Per request shape.
| Family | Pick | Why | Notes |
|---|
| Claude (Anthropic) | Opus 4.7 / Sonnet 4.6 / Haiku 4.5 / Fable 5 (Bedrock GA 2026-06-09) | Strong long-context reasoning + 1M flat-rate context on current 4.x family + tool use | Sonnet 4 + Opus 4 retire 2026-06-15 — migrate before |
| Nova 2 (Amazon) | Nova 2 Lite (GA Dec 2025), Nova 2 Pro (Preview Dec 2025) | Low/Med/High thinking modes, 1M context, code interpreter, web grounding, remote MCP tools | Pro access via Nova Forge customers |
| Nova (legacy 1.x) | Micro / Lite / Pro / Premier / Canvas (image) / Reel (video) | Cheaper than Claude at small/medium model tier; Canvas+Reel are first-party generative-media | Mature |
| Llama (Meta) | Llama 3.x (3.1 8B/70B, 3.2 11B/90B, 3.3 70B) | Open-weights; can be self-served if you outgrow Bedrock | All five sizes supported by IPR |
| Mistral / Cohere / AI21 | Mistral Large / Cohere Command / AI21 Jurassic | Multi-vendor diversification; vendor-lock mitigation | Pricing comparable to Claude family |
Thinking-level + 1M-context decision rule: if your task is multi-step reasoning over a large context, Nova 2 Pro (when GA) or Claude Sonnet 4.6 with extended thinking is the modern default. Don't pay for Opus unless your eval shows it actually wins.
3. Inference pattern. Per workload shape.
| Pattern | When | Cost class |
|---|
| On-demand | Default; bursty traffic, low volume | $/Mtok per model; no commitment |
| Provisioned Throughput (PT) | Sustained high QPS (≥several hundred RPM continuous); SLA-bound latency | Hourly $$ per model unit; commitment terms apply |
| Batch | Async jobs; 24-hr SLA acceptable | 50% off on-demand |
| Cross-Region Inference | Throughput beyond single-region limits OR data-residency-aware fan-out | Geographic (residency-bounded) vs Global (latency-optimized) — pick by compliance posture |
Decision rule: start on-demand; switch to PT only when sustained > breakeven volume; use Batch for any non-realtime workload (immediate 50% savings).
4. RAG via Knowledge Bases. When and how.
- Vector store matrix (pick by ops + scale + cost):
- OpenSearch Serverless — default; managed, auto-scale
- OpenSearch Managed Cluster (Mar 2025) — when you want capacity control / cost predictability
- Aurora PostgreSQL (pgvector) — when the rest of your data already lives in Aurora; hybrid search GA Apr 2025
- Neptune Analytics — graph + vector; for relationship-heavy retrieval
- Pinecone / MongoDB Atlas / Redis Enterprise — third-party; pick if you have an existing contract
- Binary vector support: OpenSearch Serverless only at time of writing (other stores require float32/float16) — 4× storage savings constraint if you need cross-store portability.
- Chunking strategy — Bedrock KB offers default (300 token chunks, 20% overlap), semantic, hierarchical, and custom. For domain docs, semantic + size 500-1000 is the typical starting point.
- Embedding model — Titan Embed v2 (default), Cohere Embed Multilingual; pin the model version (change = reindex).
- Defer chunking + retrieval-eval depth to
/rag-design; this skill picks the Bedrock-side store + pipeline.
5. Agent runtime — Bedrock Agents vs AgentCore vs Flows.
| Choice | When | Notes |
|---|
| AgentCore Runtime (GA 2025-10-13) | Multi-step agents with 8-hr execution windows, session isolation, MCP + A2A protocol support, framework-agnostic (CrewAI / LangGraph / LlamaIndex / Google ADK / OpenAI Agents SDK), browser/code-interpreter access | Modern default for new agents; bundles VPC / PrivateLink / CloudFormation / tagging |
| Bedrock Agents | Action-group + KB-grounded agents; managed orchestration; simpler than AgentCore for narrow use-cases | Available alongside AgentCore; check current AWS positioning before asserting legacy status |
| Bedrock Flows (GA 2024-11-22) | Visual low-code orchestration; deterministic branching; DoWhile loops added in 2025 for iterative workflows | Best for prompt-chain + tool-call workflows without code; not for emergent agentic behavior |
Decision rule: if the agent loops freely, AgentCore. If the agent follows a fixed graph, Flows. Bedrock Agents for the middle ground (managed but action-group-bounded).
6. Guardrails posture. Bedrock Guardrails surface 5 detection techniques on a single policy.
- Content filter — toxicity, hate, sexual, violence, insults, misconduct
- Denied topics — natural-language descriptions of off-policy topics
- Word filter — profanity + custom blocklist
- PII filter — built-in detectors + custom regex
- Contextual grounding check — ensures responses are grounded in provided source material; useful for RAG hallucination prevention
- Automated Reasoning (Nov 2025) — mathematical verification of responses against formal policy rules (auto-generated test Q&A). Distinct from NLI-style techniques (Llama Guard, Presidio). Worth enabling for high-stakes domains where you can express the policy formally.
- ApplyGuardrail API — invoke Guardrails on inputs/outputs from non-Bedrock models too (e.g. self-hosted Llama). Centralizes the safety posture across model providers.
- Multimodal coverage — toxicity detection on text + image.
7. Routing + cost.
- Intelligent Prompt Routing (IPR) (GA 2025-04-22) — auto-routes prompts to different models within a family by predicted complexity:
- Claude family: Haiku, Haiku 3.5, Sonnet 3.5 v1, Sonnet 3.5 v2
- Llama family: 3.1 8B, 3.1 70B, 3.2 11B, 3.2 90B, 3.3 70B
- Nova family: Nova Pro, Nova Lite
Cost-cut claims appear in blog posts but not on the GA page itself — verify the actual savings against your own traffic mix before committing.
- Prompt Caching — Claude 5-min (1.25× write) + 1-hour (2× write, GA) tiers; cache-hit read = 0.1× standard input. Stacks with Batch 50%.
- Cross-region Inference — for throughput, not cost (still on-demand pricing).
- Batch — 50% off; default for any non-realtime workload.
- Provisioned Throughput — fixed hourly cost; only justifies at sustained high RPM.
8. MLOps wiring.
- Source code: CodePipeline / GitHub Actions → upload prompts to Prompt Management (versioned) → reference by alias in app code → flip alias to roll out.
- Evaluation: Bedrock Model Evaluation for auto + human eval workflows; results lineage in CloudWatch.
- Custom Model Import: bring your own weights (Llama fine-tune / DeepSeek / Mistral) → register in Bedrock → invoke via standard API.
- Observability: CloudWatch (invocations, latency, throttles) + CloudTrail (audit log of API calls) + per-Knowledge-Base ingestion metrics + per-Agent traces.
- Cost attribution: Cost Allocation Tags on every Bedrock resource (
env, team, app); export to Cost Explorer.
9. Lock-in + exit posture.
- Portable — your prompts (text), model weights (via Custom Model Import), KB source documents (in S3), Guardrails policies (exportable JSON).
- Sticky — Bedrock-Agents action-group orchestration, Flows graph definitions, AgentCore session state, KB ingestion pipelines (chunking + embedding config), IPR routing models.
- Cross-cloud parity — for the GCP equivalent see
/vertex-ai-design; for Azure see /azure-foundry-design. Same workload runnable on all three with adapter code if you avoid the sticky surfaces above.
Output
### Bedrock Footprint: <workload-name>
**Workload type:** <chat / RAG / agent / batch-gen / fine-tune>
**QPS / batch size:** <target>
**Data residency:** <region(s)>
**Compliance:** <HIPAA / GDPR / FedRAMP / none>
**Services enabled (with reason):**
| Service | Reason / consumer | Cost class | Failure mode |
|---|---|---|---|
**Model selection:**
| Stage | Model + version (pinned ID) | Why | Thinking / context |
|---|---|---|---|
**Inference pattern:** [on-demand / PT / batch / cross-region — with rationale]
**Knowledge Bases** (if enabled): [vector store choice + binary-vector constraint + chunking + embedding model pin]
**Agent runtime** (if enabled): [AgentCore vs Bedrock Agents vs Flows — with rationale]
**Guardrails:** [content filter / denied topics / PII / grounding / Automated Reasoning — list + ApplyGuardrail scope]
**Routing + cost levers:** [IPR family / Batch / cache tier / PT terms]
**MLOps wiring:**
- Prompt Mgmt versioning + alias rollout
- Model Eval cadence
- Custom Model Import path (if any)
- CloudWatch + CloudTrail observability
**Lock-in posture:** [portable surfaces + sticky surfaces + documented exit]
**[RISK: HIGH] choices flagged:** [list, or "none"]
**Recommended ADRs:**
1. [Service-set scope for v1]
2. [Model family + tier per stage]
3. [Inference pattern + PT breakeven]
4. [KB vector-store choice]
5. [AgentCore vs Bedrock Agents vs Flows]
6. [Guardrails policy (incl. Automated Reasoning yes/no)]
7. [IPR enable / defer]
Quality bar
- Every enabled service named with consumer + cost class + failure mode — no "enabled in case"
- Model IDs pinned (legacy IDs still work but new capabilities require new IDs)
- Sonnet 4 / Opus 4 use flagged for migration before 2026-06-15 retirement
- Batch is the default for any non-realtime workload — 50% off, no excuse not to
- Knowledge Bases binary-vector constraint named when OpenSearch Serverless picked (cross-store portability blocker)
- AgentCore vs Bedrock Agents vs Flows decision recorded as an ADR — easy to retrofit wrong
- Automated Reasoning considered for high-stakes guardrails (formal-rule verification > NLI when policy is expressible formally)
- IPR savings numbers cited from a primary source (blog post URL) or marked "verify against own traffic" — the 30%-savings claim is not on the GA page
- Lock-in posture explicit — sticky surfaces named, portable surfaces named, exit path documented
- Cost-attribution tags on every resource (
env, team, app, model_id)
What this skill does NOT do
- Does NOT write Terraform / CloudFormation / CDK — design only; pair with
/terraform-review for IaC
- Does NOT design the per-agent loop — pair with
/agent-design for loop/tools/guardrails on top of AgentCore
- Does NOT do chunking + retrieval-eval depth — pair with
/rag-design for the cross-vendor RAG dimensions
- Does NOT replace
/vertex-ai-design (GCP) or /azure-foundry-design (Azure) or /sagemaker-design (AWS classical ML) — Bedrock is the AWS-GenAI surface; SageMaker is the AWS-ML surface; pick first
- Does NOT cover AWS account / org hardening — pair with
/security-audit
- Does NOT cover MCP server production — pair with
/mcp-design (Bedrock AgentCore supports MCP-protocol tools as a consumer)
Sources
Canonical docs the dated / priced claims above are sourced from. Last doc-verified against live docs 2026-07-13 (run /doc-verify on this file to re-check):