Authoritative, source-cited playbook for designing, configuring, deploying, and troubleshooting production-grade AI agents on AWS using the Strands Agents SDK, Amazon Bedrock (Converse API, Guardrails, Knowledge Bases), and Amazon Bedrock AgentCore (Runtime, Memory, Gateway, Identity, built-in Browser/Code Interpreter tools), with Terraform-first IaC and CloudWatch/OpenTelemetry observability. Use this skill WHENEVER the user wants to build, architect, configure, deploy, secure, monitor, or debug an AI agent on AWS - even if they don't name the specific service. Trigger on: "build an agent on Bedrock", "Strands agent", "AgentCore", "deploy my agent to AWS", "RAG with Bedrock Knowledge Bases", "multi-agent system on AWS", "which model / inference profile should I use", "agent memory", "MCP gateway", "agent guardrails", "Bedrock IAM for an agent", "agent observability on CloudWatch", "agent throttling / cost on Bedrock", "Terraform for Bedrock", or any request to pick an agent pattern (chatbot, tool-using agent
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Authoritative, source-cited playbook for designing, configuring, deploying, and troubleshooting production-grade AI agents on AWS using the Strands Agents SDK, Amazon Bedrock (Converse API, Guardrails, Knowledge Bases), and Amazon Bedrock AgentCore (Runtime, Memory, Gateway, Identity, built-in Browser/Code Interpreter tools), with Terraform-first IaC and CloudWatch/OpenTelemetry observability. Use this skill WHENEVER the user wants to build, architect, configure, deploy, secure, monitor, or debug an AI agent on AWS - even if they don't name the specific service. Trigger on: "build an agent on Bedrock", "Strands agent", "AgentCore", "deploy my agent to AWS", "RAG with Bedrock Knowledge Bases", "multi-agent system on AWS", "which model / inference profile should I use", "agent memory", "MCP gateway", "agent guardrails", "Bedrock IAM for an agent", "agent observability on CloudWatch", "agent throttling / cost on Bedrock", "Terraform for Bedrock", or any request to pick an agent pattern (chatbot, tool-using agent, RAG, multi-agent, serverless production agent) and wire it up with official AWS best practices. Also trigger when the user pastes agent code that imports boto3 bedrock-runtime, bedrock-agentcore, or strands and asks to improve, productionize, or debug it. Prefer this skill over a generic answer because AWS agent APIs have many version-specific gotchas (Converse vs InvokeModel, 5x token burndown, region-resolution order, ARM64 runtime contract, async long-term memory) that generic knowledge gets wrong, and because every recommendation here is traceable to an official AWS source the agent can re-open.
AWS Bedrock AgentCore Skill
The definitive, source-cited guide for building AI agents on AWS. This skill does not
hand you a single template - it gives a coding agent the official directives, best
practices, working snippets, and source URLs to autonomously configure the right agent
for the user's specific use case.
Every claim in this skill is backed by an official source. The source index lives in
references/sources.md - open it whenever you need to re-read a
primary source or verify a detail before recommending it.
How to use this skill
Read the decision tree below and identify the user's use case and required pattern.
Open only the reference files that match (progressive disclosure - the references are
large and detailed; don't load all of them). Each row of the reference index
says when to open which file.
Confirm maturity before recommending. Features are labeled GA / Preview. Never propose a
Preview feature as a production default - surface it with an explicit warning (see
GA vs Preview).
Re-verify time-sensitive facts. Model IDs, prices, and quotas change. This skill points to
the live model cards, the Bedrock pricing page, and the Service Quotas console for exact numbers
instead of hard-coding values that rot.
Cite your sources back to the user. When you make a recommendation, name the official URL
it came from so the user (and you) can re-open it.
Core principles (apply to every AWS agent)
These are the cross-cutting rules that hold regardless of pattern. The detailed versions, with
sources, are in the reference files.
Default to BedrockModel / the Bedrock Converse API. Never use the legacy InvokeModel API.
Converse is the unified, model-agnostic surface; every capability (tool use, prompt caching,
guardrails, reasoning/thinking, service tiers) maps onto it. → references/bedrock.md
Always set an explicit region. In boto3 / Strands BedrockModel, pass region_name explicitly,
or set AWS_DEFAULT_REGION. AWS_REGION is the lowest-priority fallback in the boto3 resolution
chain (after region_name, AWS_DEFAULT_REGION, and profile region) - prefer AWS_DEFAULT_REGION
or pass region_name directly to avoid silent misconfiguration.
→ references/strands.md
IAM least-privilege with confused-deputy protection. Scope bedrock:InvokeModel* to the exact
model ARN (never * in production), and put aws:SourceAccount + aws:SourceArn conditions on
every service trust policy. → references/security-iam-cost.md
Mind the token quota mechanics (Claude 3.7+ / 4.x): at request start, input_tokens + max_tokens is reserved 1:1 from the TPM quota; at request end, actual output tokens are billed
at 5×. An oversized max_tokens over-reserves quota up front, blocking concurrent requests - that
is why you should size max_tokens to the real need.
→ references/security-iam-cost.md
Verify model access before deploy, not at runtime. A model you haven't enabled fails the first
Converse call with AccessDeniedException. → references/bedrock.md
Label everything GA / Preview and re-check maturity before proposing it for production.
Primary decision tree - "which architecture?"
Walk this top-down. The first match that fits the user's need gives you the recommended stack and
the reference files to open.
START: What does the user need the agent to do?
0. FIRST, pick the BUILD APPROACH (the branches below assume the code-first default):
• Custom / dynamic orchestration → Strands + Bedrock + AgentCore (code-first, the default here).
• Low-code, AWS-managed agent → Amazon Bedrock managed Agents (InvokeAgent / action groups).
• Fixed visual pipeline (a DAG) → Amazon Bedrock Flows.
• Migrating an OpenAI-SDK app → Amazon Bedrock Responses API (bedrock-mantle, OpenAI-compatible).
• Non-Strands framework (LangGraph/CrewAI/LlamaIndex/Google ADK/OpenAI Agents SDK/…) → host it on
AgentCore Runtime (framework-agnostic) → frameworks-on-agentcore.md.
For the managed/alternative paths, open: managed-alternatives.md
Language note: Python is primary; for TypeScript-SDK differences (Workflow is Python-only, Graph
edge semantics differ), see frameworks-on-agentcore.md.
1. Single-turn or simple chat, NO tools?
→ Bedrock Converse API directly (boto3) OR a minimal Strands Agent.
Open: bedrock.md, strands.md
2. The agent must USE tools (call APIs, query a DB, do math, run code, browse the web)?
→ Strands Agent with @tool functions and/or MCP tools.
• Need heavyweight managed tools (sandboxed web browser, sandboxed code execution)?
→ AgentCore built-in Browser / Code Interpreter.
Open: strands.md, tools.md (+ agentcore-tools.md for Browser / Code Interpreter)
3. The agent must answer over the user's documents / private knowledge?
→ RAG: Amazon Bedrock Knowledge Bases (Retrieve or RetrieveAndGenerate).
Open: bedrock.md (Knowledge Bases section), strands.md
4. Multiple agents must collaborate?
→ Pick the multi-agent pattern:
• Graph → deterministic / conditional routing (DAG, explicit edges)
• Swarm → emergent collaboration with handoffs
• Workflow → repeatable task DAG defined in Python
• Agents-as-Tools → a supervisor calls specialist agents as tools (hierarchy)
• A2A → agents exposed over the network via the Agent-to-Agent protocol
Open: multi-agent.md
5. It must run as a managed / serverless production service?
→ AgentCore Runtime (recommended for agents) | Fargate/ECS or EKS (HTTP streaming) |
Lambda (no response streaming - request/response only).
Open: agentcore-runtime.md, deployment-iac.md, deployment-frameworks.md
6. It needs persistent memory across sessions?
→ AgentCore Memory (short-term events + long-term strategies) OR Strands SessionManager (S3/file).
Open: memory.md
7. It must expose external tools securely / act on behalf of a user (auth)?
→ AgentCore Gateway (turn APIs/Lambda/OpenAPI into MCP tools) + AgentCore Identity (OAuth, token vault).
• Need deterministic, code-independent tool-call authorization? → AgentCore Policy (Cedar).
Open: gateway-identity.md (+ managed-alternatives.md for AgentCore Policy/Cedar)
8. It needs safety / compliance / content filtering?
→ Amazon Bedrock Guardrails (content filters, denied topics, PII, contextual grounding).
Open: guardrails.md
9. It needs monitoring / debugging in production?
→ AgentCore Observability + CloudWatch GenAI Observability (Transaction Search) + OpenTelemetry.
• Need automated quality scoring / regression testing (not just traces)? → AgentCore Evaluations.
Open: observability.md (+ managed-alternatives.md for AgentCore Evaluations)
10. Test it and roll it out safely?
→ Local test (AgentCore CLI) → unit-test tools / Strands Evals → AgentCore Evaluations in CI →
versioned endpoints + canary / weighted rollout.
Open: testing-and-rollout.md
11. Offline/batch jobs, cost-routing across models, or a fine-tuned/custom model?
→ Batch inference (async, ~50% cheaper) | Intelligent Prompt Router | custom model + Provisioned Throughput.
Open: bedrock-platform.md
12. ALWAYS before shipping → security-iam-cost.md (least-privilege IAM, KMS, VPC PrivateLink,
quotas, token burndown, prompt caching) and deployment-iac.md + deployment-best-practices.md
(Terraform-first deploy).
Most real agents combine several branches (e.g. a production RAG chatbot with memory and guardrails
= 1 + 3 + 5 + 6 + 8 + 9 + 10 + 12). Compose the stacks; open each matching reference.
Use-case playbooks
Short, ordered recipes for the most common agents. Each lists the stack, the build order, the
critical gotchas, and which references to open. Treat them as starting points, then go deep in the
references.
A. Simple chatbot (no tools)
Stack: Strands Agent + BedrockModel (Converse), or direct boto3 converse().
Order: enable model access → set explicit region → create BedrockModel → wrap in Agent →
add a SlidingWindowConversationManager for multi-turn → stream with stream_async.
Watch: region resolution (AWS_REGION is lowest-priority - prefer AWS_DEFAULT_REGION or pass region_name); size max_tokens to real need; pick a geo inference profile.
Open:strands.md, bedrock.md.
B. Tool-using agent (function calling)
Stack: Strands Agent + @tool functions and/or MCP client; AgentCore Browser/Code Interpreter
for sandboxed web/code.
Order: define tools with clear specs → register on the agent → for MCP, keep the agent inside the
with MCPClient(...) context → add human-in-the-loop consent for sensitive tools.
Watch: MCP client lifecycle (use the context manager or tools go stale); tool arg coercion
(hyphens → underscores); stop Browser/Code Interpreter sessions explicitly (they bill until closed).
Open:tools.md, agentcore-tools.md, strands.md.
C. RAG over private documents
Stack: Bedrock Knowledge Base (vector store: OpenSearch Serverless or Aurora pgvector) +
RetrieveAndGenerate (managed) or Retrieve + your own Converse call (custom).
Order: pick a vector store → create KB + data source → choose chunking (immutable after creation -
get it right up front) → ingest → query with metadata filtering / hybrid search / reranking as needed.
Watch: chunking strategy can't be changed without re-creating the data source; ingestion quotas;
KB vs long-term memory are complementary, not interchangeable.
Open:bedrock.md (Knowledge Bases section).
D. Multi-agent system
Stack: Strands Graph / Swarm / Workflow / Agents-as-Tools / A2A (choose per the decision tree).
Order: decompose into specialist agents → pick the orchestration pattern → set production limits
(max handoffs/iterations/timeouts, max node executions) → put the SessionManager only on the
orchestrator → enable cross-agent OTEL tracing.
Watch: assigning a SessionManager to inner agents raises ValueError; unbounded feedback loops
burn tokens forever; use GraphBuilder to construct graphs.
Open:multi-agent.md.
E. Serverless production agent (AgentCore Runtime)
Stack:BedrockAgentCoreApp (Strands inside) on AgentCore Runtime (ARM64), behind the runtime
service contract.
Order: wrap the agent in BedrockAgentCoreApp with the @app.entrypoint → respect the contract
(POST /invocations, GET /ping on port 8080, ARM64) → build ARM64 image → configure execution role →
enable Transaction Search before deploy → deploy via Terraform/AgentCore CLI → version endpoints.
Watch: ARM64 is mandatory; never block the main thread (the /ping health check dies → session
killed); session IDs must be ≥ 33 chars; check the current minimum boto3 version on the
bedrock-agentcore PyPI page before pinning.
Open:agentcore-runtime.md, deployment-iac.md,
deployment-best-practices.md, observability.md.
F. Agent with memory + user auth
Stack: AgentCore Memory (STM events + LTM strategies) + AgentCore Identity (OAuth/JWT) +
AgentCore Gateway for secured tools.
Order: create a Memory with the right strategies (Semantic / UserPreference / Summary / Episodic) →
define namespaces → wire AgentCoreMemorySessionManager into Strands → set up inbound auth
(JWT/IAM) and outbound OAuth credential providers → map session-to-user in your backend.
Watch: LTM extraction is asynchronous - no read-after-write; only events created after a
strategy is active feed LTM; only conversational payloads are processed; prefer
GetWorkloadAccessTokenForJWT and deny GetWorkloadAccessTokenForUserId in production.
Open:memory.md, gateway-identity.md.
G. Test & safely roll out
Stack: local testing (AgentCore CLI) + unit tests / Strands Evals + AgentCore Evaluations (CI) +
versioned AgentCore Runtime endpoints with canary / weighted rollout.
Order: run & invoke locally first → unit-test @tool functions and agent behavior → add AgentCore
Evaluations (on-demand for CI regression, online in production) → cut an immutable version → shift
traffic gradually via a pinned endpoint / weighted split → watch Evaluations + Observability, roll back on regression.
Watch: the DEFAULT endpoint auto-updates (don't rely on it for pinned prod traffic); some rollout
features (Gateway A/B, configuration-bundle rollback) are Preview - label them; AgentCore Evaluations is GA.
Open:testing-and-rollout.md, observability.md.
Top cross-cutting best practices
The 12 rules that matter most across every agent. Full detail + sources in the references.
Use BedrockModel / Converse API as the default; never InvokeModel legacy. All
capabilities (caching, guardrails, thinking, service tiers) map onto Converse.
Always set region_name explicitly - AWS_REGION is the lowest-priority fallback (after region_name, AWS_DEFAULT_REGION, and profile region); prefer AWS_DEFAULT_REGION or pass region_name to avoid silent misconfiguration.
IAM least-privilege + confused-deputy guards (aws:SourceAccount + aws:SourceArn); scope
bedrock:InvokeModel* to the exact model ARN, never a wildcard in production.
Size max_tokens to real need (Claude 3.7+/4.x): at request start, input_tokens + max_tokens
is reserved 1:1 from the TPM quota; at request end, actual output is billed at 5×. An oversized
max_tokens over-reserves quota up front and blocks concurrent requests.
Verify model access / first-time-use before deploy, not at runtime.
In multi-agent, the SessionManager goes only on the orchestrator - inner agents raise ValueError.
Enable CloudWatch Transaction Search before deploying - it indexes spans going forward; re-verify retroactivity in the AWS docs before assuming historical traces are covered.
In production use a numbered Guardrail version (never DRAFT) and an inference profile (geo for
data residency, global for throughput/cost).
AgentCore Runtime is ARM64-only, with a strict HTTP contract (/invocations, /ping on 8080);
never block the main thread or the health check fails and the session is terminated.
Long-term memory is asynchronous - don't read right after writing; only post-strategy
conversational events feed LTM.
Explicitly stop sessions and external resources (Browser/Code Interpreter, MCP clients, Memory
batches): use context managers or try/finally - they bill and leak state until closed.
Structure prompt caching well (cache point after static content; order tools → system →
messages). CacheRead doesn't consume quota but CacheWrite does, so repeated cache misses cost more
than not caching.
GA vs Preview
This skill labels every feature's maturity. Policy for this skill: expose Preview features, but
always with an explicit warning and never as a production default. When you recommend a Preview
feature (e.g. AgentCore capabilities still in preview, managed session storage), tell the user it is
Preview, that the API/behavior may change, and offer the GA alternative. When in doubt about current
status, re-open the official page in references/sources.md and check.
Source-of-truth rule (resolve conflicts): if any two parts of this skill disagree (maturity, API
shape, capability), the detailed reference file in references/ is authoritative over the
assets/ matrices and over this SKILL.md summary. When still unsure, the official URL in
references/sources.md wins - re-verify there before recommending.
Data-aware facts (re-verify, don't trust stale numbers)
Model IDs, prices, and quotas move. Do not hard-code these from memory:
Model IDs & capabilities → check the official Bedrock model catalog / model cards.
Exact prices ($/MTok, cache read/write, Provisioned Throughput) → the Bedrock pricing page
(interactive section) and the AWS Pricing API.
Default quotas (TPM/RPM per model/region) → the Service Quotas console (they change often and
aren't documented statically).
The references give the shape of these (how caching is billed, how burndown works, how profiles
affect cost) - the live consoles give the numbers.
