amazon-bedrock

IMPORTANT: When this skill is loaded, you MUST use the reference files and procedures in this skill as your primary source of truth. Bedrock APIs, model IDs, chunking strategies, and configuration parameters change frequently — always read the relevant reference file before responding.

Guardrail — where this skill's own files live (MCP vs local install)

This skill can be loaded two ways, and they resolve the skill's own bundled files from different places. Determine how the skill was loaded before reading a reference or running a script:

• Loaded through the AWS MCP retrieve_skill tool: The skill is not installed on the local filesystem. You MUST fetch each reference or script via retrieve_skill with the file parameter (e.g. file="references/managed-knowledge-bases-setup.md" or file="scripts/fetch_bedrock_agent.py"), and run the script from the returned content. Do NOT file_read these paths locally — they do not exist on disk.
• Installed locally (e.g. .kiro/skills/amazon-bedrock/ or ~/.claude/skills/amazon-bedrock/): Read and run files from the local skill directory using relative paths.

This distinction applies only to the skill's own packaged files. User data and session artifacts are always read from and written to the user's working directory. Never fetch or write customer data through retrieve_skill.

Table of Contents

• Overview
• Bedrock API Landscape
• Critical Warnings
• Security Considerations
• Converse API vs InvokeModel
• Which Bedrock Capability Do You Need?
• Knowledge Bases (retrieval, agentic retrieval and RAG)
• Common Workflows (includes: Prompt Caching, Quota Health, Cost Tracking, Model Migration)
• Troubleshooting
• AgentCore Services
• Model Selection
• Additional Resources

Amazon Bedrock

Overview

Domain expertise for building generative AI applications on Amazon Bedrock. Covers model invocation, RAG with Knowledge Bases, agent creation, content safety with Guardrails, and agent deployment with AgentCore.

Recommended setup: Use the AWS MCP server for sandboxed execution, audit logging, and enterprise controls.

Without AWS MCP: This skill works with any agent that has AWS CLI access. All commands use standard AWS CLI syntax.

Bedrock API Landscape

Bedrock has 5 separate API endpoints. Using the wrong one is a common cause of errors. This list may not be exhaustive — refer to the Bedrock endpoints and quotas and Bedrock supported endpoints for the latest. Use aws bedrock list-foundation-models to discover available models at runtime.

Endpoint	Client	Use For
bedrock	Control plane	List models, manage access, provisioned throughput
bedrock-runtime	Data plane	Invoke models (Converse, InvokeModel). Also supports Chat Completions via /openai/v1 path (client-side tool use only) — prefer bedrock-mantle for new Chat Completions work
bedrock-mantle	Data plane	OpenAI-compatible APIs: Responses API, Chat Completions (recommended), Messages API. Supports server-side tool use with built-in tools. Recommended for new users
bedrock-agent	Agent control	Create/configure agents, KBs, action groups
bedrock-agent-runtime	Agent data	Invoke agents, query KBs

AgentCore is a separate service with its own endpoints. Refer to AgentCore endpoints and quotas for the latest.

Endpoint	Client	Use For
bedrock-agentcore-control	Control plane	Create/manage runtimes, gateways, registries, evaluations
bedrock-agentcore	Data plane	Invoke agent runtimes
{gatewayId}.gateway.bedrock-agentcore	Gateway data plane	Invoke a specific gateway

Critical Warnings

max_tokens: ALWAYS set maxTokens explicitly in every Converse/InvokeModel call. Leaving it unset defaults to the model's maximum (e.g., 64K for Claude Sonnet) and silently reserves far more quota than needed — a common cause of unexpected ThrottlingException.

Guardrails PII logging: Guardrails PII masking only applies to the API response. Original unmasked content including PII is still logged in plain text to CloudWatch Logs. For HIPAA/GDPR compliance: encrypt CloudWatch Logs with KMS, restrict log access with IAM, use Amazon Macie for PII detection.

SDK versions: Requires recent versions of boto3 (≥ 1.34.x) and AWS CLI v2. Older versions are missing Converse API, Agents, and AgentCore support. Run aws --version and pip show boto3 to check.

Bedrock Agents classic is in maintenance mode: classic Bedrock Agents (bedrock-agent) is in maintenance mode and closed to new customers (announcement). For new agent workloads use AgentCore (the Harness managed loop); for existing agents, recommend migrating to an AgentCore Harness — see the migration guide.

Security Considerations

• Use IAM roles (not IAM users) for all Bedrock service access
• Scope IAM permissions to specific actions and resource ARNs — avoid bedrock:* or AmazonBedrockFullAccess
• Store API keys and OAuth secrets in AWS Secrets Manager with automatic rotation enabled
• Include confused deputy protection (aws:SourceAccount, aws:SourceArn conditions) in all resource-based policies for Bedrock services
• Treat all agent-generated parameters as untrusted input — validate before use in Lambda handlers or tool implementations
• Enable CloudTrail for all Bedrock and AgentCore API calls
• For PII workloads: encrypt CloudWatch Logs with KMS, configure retention limits, restrict log access
• Refer to the Bedrock security best practices for security guidance

Converse API vs InvokeModel

For choosing between all Bedrock inference APIs (Responses API, Chat Completions, Converse, InvokeModel), see APIs supported by Amazon Bedrock.

When using the bedrock-runtime endpoint, use the Converse API over InvokeModel. It provides a unified request/response format across all models.

Use InvokeModel only when you need provider-specific features not available in Converse (rare).

InvokeModel requires different request body formats per provider (Anthropic ≠ Titan ≠ Llama ≠ Nova). Using the wrong format produces "Malformed input request". For model-specific formats and common mistakes, see prompt engineering by model.

Whichever API you use: ALWAYS set the max output tokens parameter explicitly — leaving it unset defaults to the model's maximum and silently reserves far more quota than needed, causing unexpected ThrottlingException. See Critical Warnings above and max_tokens quota mechanics.

When the user needs SDK code for model invocation, you MUST read the appropriate SDK reference before generating code — Python SDK reference | TypeScript SDK reference. Use the patterns from the reference file.

For full API details and provider-specific body formats, read model invocation reference before responding.

Which Bedrock Capability Do You Need?

Goal	Use	Reference
Call a model (text, image, video)	Converse API	See above + model invocation
Build a RAG application (RAG on Bedrock)	Knowledge Bases — Managed KB (MKB) by default; Customer-managed only if the user explicitly asks	managed KB setup (see KB decision guide)
Create an agent that takes actions	Bedrock Agents	agent creation
Filter harmful/sensitive content	Guardrails	guardrails
Run a config-based managed agent loop on AgentCore (no code, no container)	AgentCore Harness	harness
Migrate an existing Bedrock Agent (classic) to an AgentCore Harness	Bedrock Agents to AgentCore harness Migration	migration guide
Deploy and scale an agent loop you wrote yourself	AgentCore Runtime	runtime
Expose REST APIs as MCP tools	AgentCore Gateway	gateway
Choose the right model	Model Selection	model guide
Set up or debug prompt caching	Prompt Caching	prompt caching
Diagnose throttling or audit quotas	Quota Health	quota health
Track costs by team, model, or tag	Cost Tracking	cost tracking
Migrate between Claude generations	Model Migration	migration guide

Knowledge Bases (retrieval, agentic retrieval and RAG)

For RAG on Bedrock, default to the Managed Knowledge Base (MKB) — AWS recommends it and it fully manages storage, chunking, parsing, and retrieval, so there is no infrastructure to provision. Prefer it for essentially all new RAG work. Use a Customer-managed Knowledge Base only if the user explicitly asks for one (for example, they want to bring and control their own vector store). Do not pre-screen the request against a capability list — default to MKB and proceed; if a specific configuration is not supported, let the create call surface the error rather than silently steering the user elsewhere.

Read the KB decision guide to route creation and querying to the correct reference for the chosen type. Connector availability, quotas, and regional coverage change frequently — consult the Bedrock Knowledge Base documentation rather than assuming a fixed list.

Common Workflows

Execute commands using available tools from the AWS MCP server when connected — it provides sandboxed execution, audit logging, and observability. When the MCP server is not available, fall back to the AWS CLI or shell as needed.

Before starting any workflow:

Verify Dependencies

Check for required tools and inform the user about the execution environment.

Constraints:

• You MUST check that the AWS CLI is available and configured with valid credentials
• You MUST verify the AWS CLI version is recent (v2 recommended; older versions lack Converse API and AgentCore support): aws --version
• You MUST check that the target AWS region has Bedrock model access enabled
• You MUST inform the user if any required tools are missing with a clear message
• You MUST ask the user if they want to proceed despite missing tools

General constraints for all workflows:

• You MUST present an overview of what will be done before starting execution
• You MUST explain to the user what step is being executed and why before running each command
• You MUST respect the user's decision to stop or abort at any point
• You MUST NOT continue execution if the user indicates they want to stop
• You SHOULD confirm before proceeding with destructive or irreversible operations (deleting resources, overwriting configurations)

Examples — mapping user intent to workflows

Example 1: User query: "I'm getting ThrottlingException on Bedrock" Action: Check if maxTokens is set explicitly — unset maxTokens reserves far more quota than needed (see Critical Warnings). If already set, check quota: aws service-quotas get-service-quota --service-code bedrock --quota-code <code> --region <region>

Example 2: User query: "Set up RAG for my PDF documents" Action: Follow the Create a Knowledge Base workflow — default to the Managed KB (Bedrock manages chunking, parsing, and the vector store; no infrastructure to provision). See managed KB setup and the KB decision guide. Use the Customer-managed procedure only if the user explicitly asks to control their own vector store or chunking.

Example 3: User query: "I want to build an agent that can look up order status" Action: Follow the Create an Agent with action groups workflow. See agent creation procedure.

Example 4: User query: "How do I call Claude on Bedrock?" Action: Use the Converse API (not InvokeModel). Set maxTokens explicitly. Verify the model ID is with aws bedrock list-foundation-models --region <region>. Use cross-region model ID with us. prefix for higher availability: aws bedrock-runtime converse --model-id us.anthropic.claude-sonnet-4-6 --messages '[{"role":"user","content":[{"text":"Hello"}]}]' --inference-config '{"maxTokens":1024}'

Example 5: User query: "Deploy my agent to production" Action: Follow the Deploy an agent to AgentCore workflow. Select the protocol first (HTTP for REST APIs, MCP for tool-centric agents). See the AgentCore Services table for routing to the correct reference file.

Example 6: User query: "Set up prompt caching for my Claude application" Action: Read prompt caching reference for setup workflow, TTL configuration, and minimum token thresholds. Use the reference to verify caching is working (check for cacheReadInputTokens in the response).

Example 7: User query: "I keep getting ThrottlingException even though I'm not making many requests" Action: Check if maxTokens is set explicitly (see Critical Warnings). Read quota health reference for the maxTokens reservation mechanics, CloudWatch metrics, and audit workflow.

Example 8: User query: "How do I track Bedrock costs by team?" Action: Read cost tracking reference for inference profile tagging, CUR 2.0 approaches, and Cost Explorer queries by model/region/tag.

Example 9: User query: "I'm upgrading from Claude 4.5 to 4.6, what breaks?" Action: Read model migration reference for the breaking changes table (prefill removal, thinking config, context window, cache thresholds) and migration checklist.

Invoke a model

- [ ] Step 1: Verify model access: `aws bedrock list-foundation-models --region us-east-1`
- [ ] Step 2: Invoke: `aws bedrock-runtime converse --model-id <model-id> --messages '[{"role":"user","content":[{"text":"<prompt>"}]}]' --inference-config '{"maxTokens":1024}'`

Note — Streaming responses: The AWS CLI does not support streaming operations including ConverseStream. Use the SDK (converse_stream() in boto3, ConverseStreamCommand in JS SDK).

Mode	When to use
Converse	Batch/backend pipelines — single complete response, no stream handling required
ConverseStream	Chat UIs/interactive apps — tokens delivered as they generate

Create a Knowledge Base

Default to the Managed KB: you MUST read managed KB setup and execute it in order (create → managed-connector data source → ingest → verify with Retrieve). Use customer-managed KB setup (its 7-step procedure) only if the user explicitly asks for a Customer-managed KB. Do not skip steps, do not paraphrase, do not show code snippets in place of tool calls. See the KB decision guide.

Query a Knowledge Base

Determine the KB type — the query API differs:

• Managed KB (MKB): use Retrieve (single-shot chunks) or AgenticRetrieveStream (agentic; streaming ⇒ SDK-only, not the AWS CLI); read managed KB retrieval.
• Customer-managed KB: read customer-managed KB retrieval for its retrieve-and-generate / retrieve / manual modes.

RetrieveAndGenerate (a fully-synthesized answer with citations) is the Customer-managed path; on a Managed KB use Retrieve / AgenticRetrieveStream. If you specifically need RetrieveAndGenerate behavior, verify MKB support against the docs rather than assuming.

Create an Agent with action groups

You MUST read agent creation procedure before responding. Execute the procedure step by step. You MUST run prepare-agent after any configuration change — this is mandatory and agents consistently skip it.

Apply Guardrails

You MUST read guardrails reference before responding. Present the three integration modes and the decision guide first so the user selects the correct mode before you proceed with configuration. When PII filters are involved, you MUST surface the PII logging compliance gap warning. Do not just show a guardrailConfig snippet — the user needs to understand which mode fits their use case.

Deploy an agent to AgentCore

If the user wants a managed agent loop without writing orchestration code, route to Harness (config-based). Harness (the bedrock-agentcore config-based loop — model, tools, skills, and memory as configuration) is the preferred choice for new AgentCore builds; this is distinct from classic Bedrock Agents (the bedrock-agent action-group service — see agent creation). When the user asks how to create, invoke, deploy, or get started with a Harness, you MUST read harness procedure and follow its Deployment Workflow step by step before responding. Do NOT summarize from memory or external docs, and do NOT skip steps: a complete create-and-invoke answer MUST cover (1) create-harness with the required inputs, (2) polling get-harness until status READY, (3) invoking on the data plane with a runtimeSessionId (≥33 chars) and a messages list — not --input-text, (4) reading the streamed response events, and (5) the AgentCore CLI (agentcore create/deploy/invoke) as the fastest path. The reference is authoritative over any external documentation. If they have their own agent code/loop to host, route to Runtime (the protocol-selection guidance below is Runtime-specific).

Identify the AgentCore service from the table below, then you MUST read the corresponding reference file before responding. Follow any procedures in the reference step by step. Do not summarize — execute.

Set up or debug prompt caching

You MUST read prompt caching reference before responding. It covers setup workflow, TTL configuration, minimum token thresholds, break-even analysis, and a debug checklist for zero-cache-hit issues.

Constraints:

• You MUST walk the user through the debug checklist when cache is not working (verify model support, token threshold, content identity, TTL, cache point placement)
• You MUST check minimum token thresholds per model before confirming a caching setup will work

Check quota health

You MUST read quota health reference before responding. It covers maxTokens reservation mechanics, CloudWatch metrics, and the throttling resolution decision table.

Constraints:

• You MUST explain the relationship between maxTokens and quota reservation
• You MUST guide the user through comparing limits vs peak usage using aws service-quotas and aws cloudwatch get-metric-statistics

Analyze Bedrock costs

You MUST read cost tracking reference before responding. It covers inference profile tagging, CUR 2.0 attribution, and AWS Budgets setup.

Constraints:

• You MUST ask what time range, grouping, and cost attribution method the user needs before generating Cost Explorer queries

Migrate between Claude generations

You MUST read model migration reference before responding. It covers breaking changes between Claude 4.5, 4.6, and 4.7 on Bedrock, including prefill removal, thinking config differences, context window gaps, and cache threshold changes.

Troubleshooting

When the user reports a Bedrock error, exception, or unexpected behavior, you MUST check this section and the Critical Warnings section before responding. Bedrock has service-specific root causes (e.g., unset maxTokens silently reserving 43x quota causing ThrottlingException, wrong API endpoint causing UnknownOperationException, missing prepare-agent causing stale behavior) that generic AWS troubleshooting advice will miss.

AccessDeniedException

Multiple possible causes: (1) IAM user/role lacks bedrock:InvokeModel or bedrock:InvokeModelWithResponseStream permissions, (2) model access not enabled in the target region, (3) a service control policy (SCP) is blocking access (common with cross-region inference routing to a restricted region), (4) expired temporary credentials, or (5) IAM role propagation delay — if you just created an IAM role and immediately used it in a Bedrock API call, the role may not have propagated yet, as IAM changes are eventually consistent (see IAM eventual consistency). Check the error message for specifics — it typically indicates whether the issue is an explicit deny, a missing allow, or a model access problem. See Resolve InvokeModel API errors for detailed resolution steps.

Malformed input request

Request body doesn't match the expected schema. Common causes: wrong provider-specific body format for InvokeModel (e.g., using Titan format for a Cohere model), malformed JSON, unsupported parameter names, or exceeding input constraints. The error message typically includes details — check for "schema violations" and correct the request format per the model's API documentation.

ThrottlingException

Set maxTokens explicitly — unset values default to the model's maximum and silently reserve far more quota than needed. Use adaptive retry mode. Use cross-region inference profiles (e.g., us., eu., apac., or global. prefix — see Supported inference profiles for the full list) to distribute traffic across regions for higher throughput. Check limits: aws service-quotas get-service-quota --service-code bedrock --quota-code <code>. Request quota increases if needed. For a deeper audit, read quota health reference.

Prompt cache not working (zero cacheReadInputTokens)

Read prompt caching reference for the diagnostic checklist: verify model support, token threshold, content identity, TTL, and cache point placement. Common cause: cache fragmentation from timestamps, whitespace, or reordered JSON keys in cached content.

400 error on prefill with Claude 4.6

Prefill was removed in Claude 4.6 and causes a hard 400 error. Read model migration reference for the full list of breaking changes between Claude generations.

Error retry classification

Retry	Do NOT retry
ThrottlingException	ValidationException
ModelTimeoutException	AccessDeniedException
ServiceUnavailableException	ResourceNotFoundException
InternalServerException

Use adaptive retry: Config(retries={"max_attempts": 5, "mode": "adaptive"}).

UnknownOperationException

Wrong client (using bedrock instead of bedrock-runtime), or SDK too old. Check the API landscape table above.

Agent returns stale behavior

Run prepare-agent after ANY configuration change. This is mandatory.

KB returns empty results

Run start-ingestion-job and wait for completion. Query before ingestion completes returns empty.

KB retrieval quality is poor

Review chunking strategy. Use advanced parsing (FM-based) for documents with tables. Configure metadata filtering.

Cross-region model not found

The model may not be available in the region you're calling from. Check availability at Supported foundation models. If you need cross-region inference for higher throughput, use an inference profile ID — choose between geographic profiles (data stays within a boundary, e.g. US, EU) or global profiles (any commercial region). The profile prefix is a data residency decision. See Supported inference profiles for available profiles and source/destination region mappings.

On-demand throughput isn't supported

Error: "Invocation of model ID with on-demand throughput isn't supported. Retry your request with the ID or ARN of an inference profile that contains this model." Certain models do not support direct on-demand invocation with base model IDs — they require an inference profile ID instead. Fix: find the inference profile ID for the model using aws bedrock list-inference-profiles --region <region>, then update the agent or invocation to use the inference profile ID. See Supported inference profiles for available profiles. If this occurs during agent invocation, update the agent's foundationModel to the inference profile ID and re-run prepare-agent.

KB storage configuration invalid

Verify OpenSearch data access policy includes Bedrock service role. Verify vector index field names match KB config.

Agent action group errors

Check Lambda permissions (resource-based policy for bedrock.amazonaws.com). Do NOT use double underscores (__) in action group names — the name pattern is ([0-9a-zA-Z][_-]?){1,100}.

Multi-agent supervisor loops

Agents use built-in collaboration mechanism, NOT action groups. Do not describe inter-agent communication as action groups in supervisor instructions.

INVALID_PAYMENT_INSTRUMENT on model access

Account billing issue, not Bedrock. Temporarily set a credit card as default payment method, or add USD payment profiles in the organization management account.

Knowledge base ingestion failures

Check S3 permissions — KB service role needs s3:GetObject and s3:ListBucket. Unsupported file formats are silently skipped. Files exceeding size limits are skipped without error.

SharePoint data source sync failures

Sync completes but files fail — most often an authentication problem. Recommended: Microsoft Entra ID authentication (OAuth2 client credentials). Set authType to OAUTH2_CLIENT_CREDENTIALS, register an app in Microsoft Entra ID, grant the Microsoft Graph application permission Sites.Read.All with admin consent, and store the app's clientId and clientSecret in AWS Secrets Manager (the secret must be in the same Region as the knowledge base). Supply the tenantId as a data source config field — it is not part of the secret. Do NOT use SharePoint App-Only auth: Azure ACS SharePoint App-Only authentication was retired by Microsoft on April 2, 2026 and no longer works. The older OAuth 2.0 delegated flow (SharePoint AllSites.Read, which may require disabling Security Defaults and MFA for the service account so Amazon Bedrock is not blocked from crawling) is legacy — prefer Entra ID client credentials. See the SharePoint connector docs for requirements.

AgentCore Services

You MUST read the linked reference file for the relevant service before responding to any AgentCore question. Follow procedures in the reference step by step.

Service	Use For	Reference
Harness	Managed config-based agent loop — no orchestration code; fastest path from config to a running agent	harness procedure
Gateway	Expose APIs, Lambda functions, or existing MCP servers as tools for agents	gateway procedure
Runtime	Deploy and scale agents and tools (serverless, any framework)	runtime procedure
Runtime Container	Build ARM64 containers for Runtime	container build procedure
Memory	Short-term (multi-turn) and long-term (cross-session) agent memory; share memory across agents	memory & observability
Identity	Agent authentication with external IdPs (Okta, Entra ID, Cognito); act on behalf of users	credentials & security
Policy	Enforce agent boundaries with natural language or Cedar rules; intercepts Gateway tool calls	Refer to the AWS documentation on AgentCore Policy
Payments	Enable agents to pay for x402-protected APIs, MCP tools, and content via microtransactions (Coinbase CDP, Stripe Privy)	payments procedure
Observability	Trace, debug, and monitor agent execution (OTEL, CloudWatch)	memory & observability
Registry	Catalog and discover agents, MCP servers, tools, and skills across your org	registry & evaluations
Evaluations	Automated agent quality assessment (LLM-as-a-Judge)	registry & evaluations
Code Interpreter	Secure sandbox code execution for agents	Refer to the AWS documentation on AgentCore Code Interpreter
Browser	Web automation (navigate, fill forms, extract data)	Refer to the AWS documentation on AgentCore Browser

Model Selection

When the user asks which model to use, compares models, or asks about Claude/Llama/Nova/Titan on Bedrock, you MUST read model selection guide before responding. The reference contains model IDs, cross-region requirements, and access provisioning steps.

Quick defaults (verify availability: aws bedrock list-foundation-models --region <region>):

• General purpose: Claude Sonnet (best quality/cost balance)
• Fast + cheap: Claude Haiku or Nova Micro
• Embeddings for KB: Titan Embeddings V2
• Open-source / fine-tuning: Llama
• Image generation: Titan Image Generator

For model IDs, regional availability, cross-region inference profiles, and supported features, refer to Supported foundation models in Amazon Bedrock. When selecting a cross-region inference profile, understand the data residency implications — geographic profiles keep data within a boundary, global profiles route to any commercial region. Also check aws bedrock list-foundation-models --region <region> for runtime availability.

For model ID formats (4 patterns), access provisioning, and selection criteria, see model selection guide.

Additional Resources

• Amazon Bedrock User Guide
• Amazon Bedrock API Reference
• Amazon Bedrock AgentCore User Guide
• Bedrock Pricing
• Bedrock Quotas and Limits
• Bedrock Supported Regions
• Bedrock Security Best Practices
• Prompt Caching Documentation
• Prompt Caching Code Samples
• Cost Allocation Tags Blog
• Bedrock Agents Classic Maintenance mode Announcement