---
title: Generative AI on AWS — Production-Ready LLM Apps in Weeks
description: Generative AI strategy and delivery on AWS — use-case selection, Bedrock + SageMaker architecture, governance, evaluations, and production rollout across the AWS AI stack.
url: https://www.joinhorizon.co/services/generative-ai-on-aws/
category: genai
updated: 2026-10-09
---

# Generative AI on AWS — Production-Ready LLM Apps in Weeks

> 88% of AI pilots never reach production — they stall on governance, cost spirals, and team alignment. We get yours to ship: RAG pipelines, agents, and Bedrock applications hardened with cost guardrails, evaluations, and security controls from sprint one.

## What is Generative AI on AWS?

Generative AI on AWS is the set of managed services that lets organizations build, deploy, and operate large language model (LLM) and foundation-model applications without managing GPU infrastructure. The stack centers on Amazon Bedrock for foundation models (Claude Fable 5.1 for long-horizon Covered Model jobs, Claude Opus 5.5 with ZDR default and 1M context, Claude Opus 5 as the prior Opus pin, Claude Sonnet 5, Claude Haiku 5.5 (Haiku 4.5 remains the prior pin), Amazon Nova 2 Lite, OpenAI GPT-6 Astra / Sol / Luna with GPT-5.6 as the prior pins, xAI Grok 4.6 with US Geo / Global CRIS, Grok 4.3, Llama, Mistral Large 3, Cohere, and marketplace models including GLM 5.3 (October 5, 2026 — eligible customers, US/Global CRIS, 1M context; GLM 5 remains the prior pin) and MiniMax M2.5), Amazon SageMaker AI / Unified Studio for custom training and hosting, and Amazon Q / Quick Suite for turnkey assistants — all integrated with VPC endpoints, IAM, KMS, and CloudTrail for data residency and audit-ready compliance.

## Related Case Studies

See how we've deployed production GenAI systems that deliver measurable business outcomes:

- **[Amazon Q for Developers: Accelerating Developer Productivity](https://www.joinhorizon.co/case-study/amazonq/)** — Achieved 100% adoption in 44 days with 30-50% faster code development and 35% fewer post-release defects at TargetBay.

---

## The AWS Generative AI Stack

AWS provides the most complete enterprise-ready stack for generative AI — from managed model access to data pipelines, vector search, orchestration, and guardrails.

### Bedrock vs SageMaker vs Amazon Q — Decision Matrix

| If your use case is…                                                   | Choose                                               | Why                                                                                        |
| ---------------------------------------------------------------------- | ---------------------------------------------------- | ------------------------------------------------------------------------------------------ |
| Customer-facing chatbot, RAG over docs, summarization, classification  | **Amazon Bedrock**                                   | Serverless foundation-model APIs, Knowledge Bases, Guardrails — fastest path to production |
| AI agent that calls tools/APIs across multiple steps                   | **Bedrock AgentCore**                                | Agents Classic in maintenance — Runtime, Harness, Gateway for net-new builds               |
| Custom-trained model on your proprietary data (churn, fraud, forecast) | **Amazon SageMaker**                                 | Full ML lifecycle: training, tuning, hosting, monitoring                                   |
| Fine-tuning a foundation model on domain data                          | **SageMaker JumpStart** or **Bedrock fine-tuning**   | Bedrock for managed simplicity; SageMaker for control                                      |
| Internal employee Q&A across SharePoint/Confluence/Salesforce          | **Amazon Quick Suite**                               | Net-new path after Q Business maintenance (July 30, 2026); ACL-aware Quick Index           |
| AI coding assistant in IDE/CLI                                         | **Amazon Q Developer**                               | IDE-native, /dev agent, security scans, code transformation                                |
| Conversational analytics on dashboards                                 | **Amazon Q for QuickSight**                          | Natural-language BI on existing QuickSight datasets                                        |
| Air-gapped, regulated workload that cannot use API-based models        | **SageMaker (private VPC)** with open-source weights | Full data isolation, customer-managed inference                                            |
| Knowledge base over very large vector corpus (100M+ vectors)           | **S3 Vectors** + Bedrock                             | Native S3 vector storage, ~10× cost reduction vs OpenSearch for cold archives              |

The key services:

### Amazon Bedrock

Amazon Bedrock is the starting point for most enterprise GenAI on AWS. It provides serverless access to foundation models from Anthropic (Claude), Meta (Llama), Mistral, Cohere, OpenAI, xAI, and Amazon (Nova, plus Titan embeddings) through a single API — without provisioning or managing any model infrastructure.

Bedrock includes:

- **Model invocation** — Single API for text generation, embeddings, image generation across all providers
- **Bedrock Knowledge Bases** — Managed RAG infrastructure with automatic chunking, embedding, and vector storage
- **Bedrock Agents** — Orchestration framework for multi-step AI agents that use tools and take actions
- **Bedrock Guardrails** — Content filtering, PII detection, topic restrictions, and grounding checks
- **Model evaluation** — Side-by-side comparison of models against your data before committing
- **Bedrock Prompt Caching** — Cuts input token costs 70–90% and latency 60–85% on workloads with stable system prompts or repeated knowledge-base context
- **Bedrock Marketplace & Custom Model Import** — 100+ models including third-party and your own fine-tuned weights, deployed on managed Bedrock infrastructure
- **Bedrock Flows & AgentCore** — Visual workflow builder plus a durable agent runtime. AgentCore Managed Harness (April 2026) handles deployment, scaling, and security for agentic workloads; Policy controls (GA March 2026) verify what actions agents can take before they reach tools or data
- **Bedrock Managed Agents (OpenAI)** — Limited-preview managed agents powered by OpenAI frontier models, combining GPT-5-class reasoning with AWS infrastructure and governance

### Amazon SageMaker

SageMaker is the platform for teams that need to go beyond off-the-shelf foundation models:

- **Fine-tuning** — Domain-specific customization of foundation models on your data
- **Model hosting** — Deploy custom or fine-tuned models with auto-scaling endpoints
- **SageMaker Pipelines** — Automated ML workflows for training, evaluation, and deployment
- **Feature Store** — Centralized feature management for ML applications
- **SageMaker Unified Studio** — Single workspace for data preparation, model development, and GenAI app building with shared governance

### Amazon Q

Amazon Q extends generative AI capabilities to specific AWS use cases:

- **Amazon Q Business** — Enterprise assistant connected to your internal knowledge base (SharePoint, Confluence, Salesforce, S3)
- **Amazon Q Developer** — AI coding assistant for AWS development tasks, integrated into IDEs and the CLI
- **Amazon Q for QuickSight** — Natural language interface for BI dashboards
- **Amazon Q Apps** — User-built workflows on top of Q Business (no-code, ACL-aware)
- **[Amazon Q in Connect](https://www.joinhorizon.co/services/amazon-connect/)** — Real-time agent assistance for contact centers on Amazon Connect

## What We Build

### Internal Knowledge Assistants

Enterprise chatbots that answer questions from your internal documentation, Confluence, SharePoint, Slack archives, and database records. Powered by Bedrock Knowledge Bases with Kendra or OpenSearch as the retrieval layer.

**Example:** A healthcare company's internal assistant answering clinical protocol questions from 50,000 pages of documentation — with citations and source links, deployed with HIPAA-compliant architecture on Bedrock.

### Document Intelligence

Applications that read, extract, classify, and summarize large volumes of documents — contracts, medical records, financial reports, regulatory filings.

**Components:** Textract for extraction, Bedrock Claude for summarization and classification, S3 for storage, Lambda for orchestration.

**Example:** A fintech operations team automating SOC 2 evidence collection from 12,000 quarterly documents — Textract → Bedrock classification → S3 evidence locker, audit log every step. Replaced 3 FTE-weeks of manual review per quarter.

### AI Customer Support

Customer-facing support automation that handles common inquiries, escalates complex cases, and provides agents with suggested responses — grounded in your product documentation and customer history.

### Code Generation & Review Workflows

Developer tooling that accelerates code review, generates boilerplate, writes tests, and explains complex codebases — integrated with your CI/CD pipeline and built on Amazon Q Developer or Bedrock.

### Predictive Analytics with GenAI

Combining SageMaker for predictive models with Bedrock for natural language explanation of predictions — enabling business users to understand model outputs without data science expertise.

**Example:** A retail operations team running a SageMaker demand-forecast model for 800 SKUs, with Bedrock generating plain-language explanations of week-over-week shifts ("inventory at risk at 12 stores due to promo cannibalization") for store managers who do not read statistical output.

## Our GenAI Delivery Process

### Phase 1: Discovery (Week 1)

- Use case scoping and feasibility assessment
- Data audit — what private data exists, in what format, and how it must be protected
- Architecture selection — Bedrock vs. SageMaker, RAG vs. fine-tuning, vector store selection
- Compliance requirements mapping (HIPAA, SOC 2, PCI DSS if applicable)

### Phase 2: Prototype (Weeks 2–3)

- Core application built with real data
- Model selection and evaluation
- RAG pipeline configuration and retrieval quality testing
- Initial guardrails implementation

### Phase 3: Productionize (Weeks 4–8)

- Authentication and authorization integration
- Observability (CloudWatch metrics, request/response logging)
- Cost controls and model invocation budget alerts
- Guardrails hardening and adversarial testing
- Load testing and latency optimization
- CI/CD pipeline for model prompt versioning and deployment

### Phase 4: Monitor & Improve

- Response quality monitoring
- Retrieval relevance tracking
- Model version upgrades as new foundation models release
- Continuous improvement based on user feedback

## Security & Governance

Enterprise generative AI requires more than just good prompts:

- **Data isolation** — All components deployed within your VPC. No data leaves your AWS environment.
- **Model access control** — IAM policies restrict which roles and services can invoke models
- **Audit logging** — Every model invocation logged to CloudTrail with user identity and request context
- **Guardrails** — Bedrock Guardrails for content filtering, PII protection, and topic restrictions
- **Prompt injection protection** — Input validation and system prompt hardening
- **Cost guardrails** — Per-model and per-user invocation budgets with alerts

For deep-dive guidance on specific Bedrock capabilities, see our [Amazon Bedrock Consulting](https://www.joinhorizon.co/services/aws-bedrock/) service. For machine learning beyond foundation models, see [AWS SageMaker Services](https://www.joinhorizon.co/services/aws-sagemaker/).

## Further Reading

- [Why AWS Bedrock Is the Fastest Path to Enterprise GenAI](https://www.joinhorizon.co/blog/why-aws-bedrock-is-the-fastest-path-to-enterprise-genai/) — Architecture overview
- [Bedrock Cost Optimization: Token Budgets and Model Selection](https://www.joinhorizon.co/blog/aws-bedrock-cost-optimization-token-budgets-model-selection/) — Stop inference bill surprises
- [Fine-Tuning vs RAG on Bedrock: When to Use Which](https://www.joinhorizon.co/blog/fine-tuning-vs-rag-bedrock-when-to-use/) — Decision framework
- [HIPAA-Compliant AI on AWS Bedrock](https://www.joinhorizon.co/blog/hipaa-compliant-ai-aws-bedrock/) — Regulated workload pattern
- [Multi-Agent Supervisor Pattern on Bedrock](https://www.joinhorizon.co/blog/aws-bedrock-multi-agent-supervisor-pattern/) — Production agent architecture
- [Amazon Bedrock AgentCore in Production](https://www.joinhorizon.co/blog/amazon-bedrock-agentcore-production/) — Durable agent runtime
- [EU AI Act Compliance on Bedrock and SageMaker](https://www.joinhorizon.co/blog/eu-ai-act-compliance-aws-bedrock-sagemaker/) — Regulatory readiness
- [eCommerce AI Agents Series](https://www.joinhorizon.co/resources/ecommerce-ai-agents/) — 64 parts on agents for retail: tools, autonomy, and human approval gates
- [AWS AI Agent Architecture for eCommerce](https://www.joinhorizon.co/blog/aws-ai-agents-for-ecommerce-factualminds-2026/) — CTO brief: Harness vs Runtime, Gateway, Cedar, Identity
- [Human-in-the-Loop AI Agents](https://www.joinhorizon.co/blog/human-in-the-loop-ai-agents-ecommerce-2026/) — Where approval sits when an agent can move money

[Book a Free GenAI Discovery Call →](https://www.joinhorizon.co/contact-us/)

## Features

### Amazon Bedrock Applications
Production LLM applications on Amazon Bedrock — RAG pipelines, conversational AI, document intelligence, and multi-model workflows across Claude Opus 5.5 (September 22, 2026 — ZDR default; thinking always on) and Sonnet 5.5 (September 28, 2026) for agentic work, Claude Haiku 5.5 (October 7, 2026 — 1M context; Haiku 4.5 remains the prior pin) and Nova 2 Lite for high-volume tasks, Claude Fable 5.1 (September 1, 2026 — Covered Model) for extended autonomous jobs after legal sign-off, GPT-6.1 Sol (September 29, 2026), GPT-6 Luna, and GPT-6 Astra for OpenAI lanes, and Grok 4.7 (September 28, 2026) for Converse-native xAI, Llama for fine-tuning — all without GPU infrastructure.

### RAG Pipeline Architecture
Retrieval-Augmented Generation systems that connect foundation models to your private knowledge base — documents, databases, and APIs — for grounded, accurate responses.

### AI Agents & Automation
Net-new production agents on Bedrock AgentCore — Runtime, Memory (FGAC via Gateway Cedar as of August 28, 2026), Gateway, Identity, Observability, Managed Harness (GA June 17, 2026), and optional Payments (GA August 18, 2026 — session-budgeted agent-to-API spend, not checkout). Existing Bedrock Agents Classic deployments continue to operate; Agents Classic enters maintenance for new customers July 30, 2026. Multi-agent supervisor patterns and tool-rich workflows integrate with Lambda, Step Functions, and your existing AWS services.

### Fine-Tuning & Model Customization
Domain-specific model customization using your proprietary data — improving accuracy for specialized terminology, classification tasks, and domain-specific generation.

### GenAI Security & Guardrails
Bedrock Guardrails plus Automated Reasoning Checks for math and logic-grounded validation — preventing prompt injection, PII leakage, and hallucinations on regulated content (medical, financial, legal). Custom filters for domain-specific risk.

### Prompt Caching & Cross-Region Inference
Bedrock Prompt Caching cuts repeated-context input tokens by 70–90% on RAG workloads and trims latency 60–85% on cache hits. Cross-region inference profiles route around capacity limits without code changes.

### ML Platform Engineering
SageMaker pipelines, model registries, and MLOps infrastructure for teams that need to train, evaluate, and deploy custom models at scale.

## Why FactualMinds

### Production Focus
We build systems that run in production — not demos. Every engagement includes observability, error handling, and cost controls.

### AWS GenAI Stack Expertise
20+ Bedrock systems shipped to production. We have debugged the failure modes — retrieval drift, agent loops, evaluation regression, cost runaway — so your team does not have to.

### Security-First Architecture
All GenAI builds include data privacy controls, model access governance, and compliance considerations from day one.

### Prototype to Production in Weeks
Focused prototype in 2 weeks. Production-hardened system in 6–10 weeks. Fixed scope, locked price, your team owns it after — no consulting handoff.

### Cost Guardrails from Day One
We set hard monthly spend limits and per-feature token budgets before the first user hits production. Cost monitoring dashboards are delivered in sprint one — no inference bill surprises at scale.

### Evaluation-Driven Development
We build a golden test dataset during development and run automated evaluations against it on every deployment. You receive a measurable quality baseline before launch — not a demo.

### Minimal Abstraction, Debuggable Code
We favor explicit control loops, string handling, and API calls you can read and own — not framework-heavy stacks that hide failure modes behind five layers of abstraction. Your team debugs production issues without calling us.

## FAQ

### What is generative AI on AWS?
Generative AI on AWS refers to building LLM-powered applications using AWS managed services — primarily Amazon Bedrock (for accessing foundation models from Anthropic, Meta, Mistral, and Amazon without managing infrastructure) and Amazon SageMaker (for training, fine-tuning, and hosting custom models). AWS provides a complete stack for enterprise generative AI: managed model access, vector databases (OpenSearch, RDS pgvector), retrieval pipelines, orchestration (Step Functions, Bedrock AgentCore), guardrails, and security controls that keep your data within your AWS environment. Workforce AI net-new buyers after July 30, 2026 should evaluate Amazon Quick Suite instead of Q Business.

### Why build generative AI on AWS instead of using a third-party API directly?
Direct third-party LLM APIs (OpenAI, Anthropic API) send your data to the model provider. For enterprise applications — especially in healthcare, fintech, or legal — this creates data residency, privacy, and compliance issues. Amazon Bedrock provides access to the same underlying models (including Claude from Anthropic) through an API that stays within your AWS environment. Your data never leaves your AWS account. You also get consolidated billing, IAM-based access control, CloudTrail logging of all model invocations, and integration with your existing VPCs and security controls.

### What is a RAG pipeline and do I need one?
RAG (Retrieval-Augmented Generation) is an architecture where a foundation model is given relevant context retrieved from your private knowledge base before generating a response. Without RAG, the model only knows what was in its training data — it cannot answer questions about your products, policies, customer history, or any other proprietary information. With RAG, you retrieve relevant documents from your knowledge base and include them in the model prompt, enabling accurate, grounded responses based on your data. Most enterprise GenAI applications — internal assistants, customer support bots, document Q&A — require RAG.

### How long does it take to build a production generative AI application on AWS?
A focused prototype (demonstrating core functionality with real data) typically takes 2–3 weeks. Production-hardening — adding error handling, observability, guardrails, cost controls, and user authentication — adds another 4–6 weeks. More complex systems with fine-tuning, multiple agents, or integration with many data sources take longer. We recommend starting with a focused prototype on the highest-value use case, validating it with real users, then expanding.

### Which foundation model should I use on Bedrock?
The right model depends on your use case, latency requirements, and cost tolerance. Claude Opus 5.5 (September 22, 2026 — ZDR default on Bedrock, adaptive thinking always on, 1M context) is the current Opus workhorse; Claude Opus 5 (July 24, 2026) stays the prior pin until you benchmark it. Claude Sonnet 5.5 (September 28, 2026 — 1M context, adaptive thinking on by default, also in AWS GovCloud (US)) is the balanced agentic default; Claude Sonnet 5 (June 30, 2026) stays the prior pin until you benchmark 5.5. Claude Haiku 5.5 (October 7, 2026 — 1M context, model ID anthropic.claude-haiku-5-5, Global CRIS global.anthropic.claude-haiku-5-5, Marketplace-billed) is the current volume Haiku; Claude Haiku 4.5 (October 16, 2025) stays the prior pin. Claude Fable 5.1 (September 1, 2026) is the current Mythos-class GA path for multi-hour autonomous jobs — it is a Covered Model (30-day retention / aws_review; not ZDR by default; EFS rolls out fall 2026). OpenAI GPT-6 Luna, GPT-6.1 Sol, and GPT-6 Astra cover volume, recurring complex work, and the hardest end-to-end jobs (GPT-6.1 Sol GA September 29, 2026 — OpenAI reports near-Astra results at Sol pricing, 1M context; Ultrafast October 8, 2026; Global CRIS global.openai.gpt-6.1-sol on the model card; GPT-6 Sol and Luna GA September 22, 2026, up to 1M context, with GPT-6 Sol as the prior Sol pin; Astra GA September 8, 2026). GPT-5.6 Luna/Terra/Sol remain prior pins (Luna −80% / Terra −20% July 30, 2026; 1M context August 3, 2026). Amazon Nova 2 Lite (December 2, 2025, 1M context) is the current everyday Nova; Nova Micro remains the cheapest published text lane. Grok 4.7 (September 28, 2026) is the current Converse-native xAI model — 500K context, four reasoning-effort levels, US Geo and Global CRIS; Grok 4.6 is the prior pin and Grok 4.3 remains the cheaper Mantle-only long-context option. We benchmark candidates against your data before committing.

### How do you handle security and data privacy for generative AI?
Security is built into every GenAI engagement from day one. We deploy all components within your VPC — no data traverses the public internet. Amazon Bedrock model providers never see your data. We implement Bedrock Guardrails to prevent PII leakage, prompt injection, and harmful content. All model invocations are logged to CloudTrail. IAM policies restrict model access to authorized roles. For regulated industries (healthcare, fintech), we ensure the architecture meets applicable compliance requirements (HIPAA, SOC 2, PCI DSS) before deployment.

### What is the difference between Amazon Bedrock and Amazon SageMaker?
Bedrock gives you managed access to foundation models via API; SageMaker is the platform for training, fine-tuning, and hosting custom models. Most enterprise GenAI uses both — Bedrock for the LLM, SageMaker for proprietary embeddings or predictive models. For a side-by-side feature matrix and pricing comparison, see our [Amazon Bedrock service page](https://www.joinhorizon.co/services/aws-bedrock/).

### Can you help us evaluate whether generative AI is the right fit for our use case?
Yes — our GenAI Discovery Call is specifically designed for this. We will assess your use case against four criteria: data availability (do you have the private data to power the application), business value (is the expected ROI clear), technical feasibility (can the use case be reliably solved with current LLM capabilities), and organizational readiness (does your team have the context to validate and iterate on the system). Not every use case is a good fit for GenAI, and we will tell you honestly if yours is not.

### How do you prevent inference costs from spiraling in production?
We implement hard budget limits at the AWS Bedrock account level, per-feature token budgets, and cost monitoring dashboards from sprint one. Most clients know their per-query cost within two weeks of going live — before they have committed to scale. Automated alerts trigger before spend thresholds are reached, not after.

### What are Amazon Nova models and when should we use them instead of Claude?
Amazon Nova is Amazon's own foundation model family available on Bedrock: Nova Micro (text-only, ultra-low latency, ~$0.04/1M input tokens), Nova Lite (multimodal — text, image, and video input, strong price/performance balance), and Nova Pro (highest Nova capability for complex reasoning and vision tasks). Nova Micro costs far less per token than Claude Sonnet 5.5. For high-volume tasks — classification, extraction, summarization, routing, and structured output generation — Nova delivers strong results at a fraction of the Claude cost. We benchmark both models against your specific use case and data before recommending, and many production systems we build use Nova for volume workloads and Claude Sonnet 5.5 for reasoning-intensive tasks.

### How do you measure whether the AI is actually working?
We establish a golden dataset during development — a curated set of test cases drawn from your real user queries — and run automated evaluations against it on every deployment. This gives you a quality number, not just a demo. We also set up hallucination scoring and output validation so your team can monitor quality in production, not just at launch.

### We already have a Bedrock prototype but it is not making it to production. Can you help?
Yes — this is the most common engagement we run. We start with a 1-week diagnostic: review your current architecture, retrieval quality, evaluation gap, cost trajectory, and security posture. You get a written punch list of what is blocking production with effort estimates against each item. From there we either embed with your team to ship the fixes (typical 4–6 weeks) or hand back the punch list for your team to execute. No vendor lock-in.

### How do you handle hallucinations on regulated content like medical or financial advice?
For regulated workloads we layer three controls: Bedrock Guardrails for content and PII filtering, Bedrock Automated Reasoning Checks for math and logic validation against a defined policy, and golden-dataset evaluations on every deploy. For high-stakes outputs we add human-in-the-loop review and citation requirements — the model must surface the source document for every claim. We have shipped this pattern under HIPAA and SOC 2 audit; the architecture pattern is documented in our [HIPAA-compliant AI on Bedrock guide](https://www.joinhorizon.co/blog/hipaa-compliant-ai-aws-bedrock/).

### We are using OpenAI directly today. What does migrating to Bedrock involve?
Two-week migration is typical for most applications. The work: switch the SDK from OpenAI to Bedrock (Anthropic Claude is the closest behavioral match for GPT-4-class workloads, Nova for cheaper-than-GPT-4o-mini volume tasks), rewrite prompts where needed (Claude responds slightly differently to system instructions), wire VPC endpoints for data residency, and run an evaluation against a golden dataset to confirm parity. We document the cost delta — typically 30–60% lower for comparable quality once Prompt Caching is enabled.

### How do you make sure my team can run this after you leave?
Knowledge transfer is built into every engagement, not an afterthought. Your engineers pair on every PR, our weekly demos are recorded with architecture rationale, the runbook is written in your wiki (not ours), and we run a 2-hour handoff session before invoice close. We have references from clients who ran their Bedrock systems in production for 12+ months without re-engaging us — that is the outcome we optimize for, not retainer hours.

### We hired a GenAI consultant before and got a strategy deck — or code we cannot maintain. How are you different?
We have heard that story often: impressive slides, then ghosting when it was time to build — or a handoff of spaghetti with no tests, hardcoded keys, and no docs. Every engagement here is fixed scope and locked price with production deliverables: evals, cost guardrails, observability, and code your team can read. We embed with your engineers; we do not disappear after the deck. If a prior engagement left you in pilot purgatory, our 1-week diagnostic produces a punch list you can execute with or without us.

### Do you rely on heavy LLM frameworks, or can we own the code?
We build with minimal abstraction — explicit prompts, control loops, and Bedrock/SageMaker APIs — so you can debug, monitor, and change behavior without fighting a framework. Orchestration stays in code you own (Step Functions, AgentCore, or a thin application layer). Frameworks are fine for demos; production needs rails you can inspect. Your team should be able to ship a prompt change without waiting on us.

---

*Source: https://www.joinhorizon.co/services/generative-ai-on-aws/*