Agent Sketch: LLM-Powered Creative Workflow & Design Automation on AWS

Executive Summary 

To meet the growing demand for scalable and efficient visual content creation in real time with quick prompts, we developed a cloud-native, AI-powered design automation platform Agent Sketch which is deployed on AWS. This solution leverages intelligent agents to automate the generation of logos, icons, and other creative assets, enabling rapid iteration and consistent branding across multiple channels.  

Built on Amazon EKS for container orchestration and Amazon Bedrock to access AI models, the platform is optimized for high performance, modularity, and scalability. Visual assets are stored securely in Amazon S3, while agent configurations, workflow definitions, and user data are managed using Amazon DynamoDB, providing low-latency access and flexible schema support for real-time interactions.

The architecture also incorporates real-time monitoring and observability using Amazon CloudWatch and supports seamless API integration for embedding design agents into external systems and applications. This event-driven, serverless-compatible design enables fast, scalable deployment with minimal operational overhead.

By implementing this solution, organizations can reduce manual design efforts by over 60%, accelerate time-to-market for digital campaigns, and improve brand consistency through automated asset generation. Agent Sketch on AWS provides a flexible, cost-efficient, and highly extensible platform that democratizes creative automation and empowers businesses to innovate at scale. 

Customer Challenge 

Customer Information 

Customer: [Customer Name] 
Industry: [Customer's Industry] 
Location: [Customer's Primary Location] 
Company Size: [Number of employees or relevant size metric] 

Business Challenges 

Enterprises with growing digital presence face significant challenges in scaling creative production and maintaining brand consistency across channels. Design teams are often overwhelmed by the volume of requests for visual assets, ranging from logos and icons to campaign creatives, leading to bottlenecks, delays, and inconsistent branding across regions, products, and platforms. 

Much of the creative process remains manual, time-consuming, and non-repeatable, with designers relying on traditional tools and isolated workflows. Business users lack self-service tools to generate on-brand assets, forcing over-reliance on design teams and resulting in long turnaround times. This hinders the speed of marketing campaigns, product launches, and customer engagement initiatives. 

Without centralized asset generation or control, organizations struggle with duplicate work, brand inconsistency, and limited visibility into asset usage. Design assets are scattered across storage systems with little to no versioning, metadata, or governance. AI-based design initiatives often lack auditable model outputs, rollback capabilities, or integration with compliance workflows. 

Additionally, as teams explore AI-driven creativity, they face challenges in scaling AI models, managing configurations, and ensuring recovery in case of failures, especially when models are tightly coupled with complex, fast-changing creative requirements. 

With increased demand for personalized content, faster go-to-market timelines, and cost-effective scaling of design operations, businesses require a cloud-native, AI-powered platform that can automate asset generation, enforce brand standards, and support resilient and observable design workflows exactly what Agent Sketch delivers. 

Technical Challenges 

Building an AI-powered design automation platform like Agent Sketch from the ground up presented a range of complex technical challenges—particularly in enabling scalable, secure, and explainable creative generation without relying on pre-existing infrastructure. 

First, the platform needed to support high-performance foundation models in a cloud-native environment while maintaining low latency and real-time responsiveness. Traditional hosting approaches lacked the elasticity and orchestration required to manage bursty workloads and concurrent user sessions. To address this, the system was architected using Amazon Bedrock to access pre-trained foundation models such as Stable Diffusion, and Amazon EKS to manage containerized agent services that orchestrate prompt generation, post-processing, and workflow execution. 

Secondly, there was no out-of-the-box system for managing design agent configurations, workflows, and prompt metadata in a traceable, version-controlled, and auditable manner. Since Bedrock abstracts away the need to manage model artifacts directly, the focus shifted toward tracking inputs, outputs, and contextual parameters for every model invocation. This was achieved by designing a custom asset and metadata management layer on top of Amazon S3 (with versioning enabled) and structured configuration storage using Amazon DynamoDB. This ensured traceability of generated content and reproducibility of creative workflows, even without managing model infrastructure directly  

Another critical challenge was the need for modular, reusable agent logic. Agent behavior had to be configurable by non-technical users while remaining executable in real-time pipelines. This demanded a declarative, event-driven architecture that could support chaining of AI-based operations (e.g., generating, styling, approving assets) while remaining stateless, fault-tolerant, and observable. 

Finally, the platform had to meet enterprise-grade standards for security, compliance, and explainability. With AI-generated assets used in marketing and public-facing materials, it was essential to build controls for access management, audit logging, model output traceability, and content versioning—none of which are natively available in standard design tools or general-purpose ML platforms. 

Together, these challenges necessitated the design of a fully cloud-native, modular, and extensible architecture, capable of supporting creative agents as services, orchestrating asset pipelines, and delivering scalable AI-driven design automation with reliability and trust. 

Partner Solution 

Solution Overview

Xenonstack implemented a modular, agent-based creative automation platform designed to address the scalability, consistency, and governance challenges of enterprise design workflows. Built on a cloud-native architecture and deployed using Amazon EKS, the solution orchestrates intelligent design agents for generating, transforming, reviewing, and managing digital assets such as logos, icons, and branded visuals. 

AI models accessed through Amazon Bedrock for text-to-image generation. 

Agent logic and workflows are defined declaratively, allowing business teams to configure asset generation parameters, brand constraints, and review rules without modifying code.  

A key capability of the platform is its LLM-driven design orchestration, where agents leverage models to generate creative suggestions, textual prompts, or brand-aligned asset variations. This enables a human-in-the-loop experience where designers and marketers can rapidly iterate and approve visuals with AI assistance—improving productivity and brand compliance. 

To enhance personalization and maintain creative continuity, Agent Sketch integrates a dual-layer context management system: 

• Short-Term Context using Amazon ElastiCache (Redis): Maintains session-level preferences, recently used styles, and dynamic prompt refinements for immediate reuse across tasks. 

• Long-Term Memory & Relationships using Amazon Neptune: Stores a persistent knowledge graph of user-brand interactions, approved asset lineage, campaign mappings, and design rules—enabling smarter prompt generation, enforcement of branding logic, and historical recall. 

This context engineering approach transforms Agent Sketch from a static asset generator into a memory-aware creative assistant that understands brand history, user intent, and campaign context. 

Security and scalability are built into the platform from the ground up, with IAM-based access control, private VPC endpoints, and GitOps-style CI/CD pipelines enabling seamless deployment across dev, staging, and production environments. Amazon CloudWatch is used for end-to-end observability of agent performance, asset generation latency, and model inference health, while output metrics feed into dashboards for creative ops teams to monitor performance and content readiness. 

By centralizing creative logic, automating asset workflows, and enabling intelligent design agents, Agent Sketch delivers a scalable, auditable, and extensible platform that empowers both design and marketing teams to accelerate production, maintain brand integrity, and innovate with AI at scale. 

AWS Services Used 

• Amazon EKS: Container orchestration for managing microservices 

• Amazon Bedrock: Provide access to models to generate images from text. 

• Amazon S3: Primary storage for generated design assets, model artifacts, logs, and agent outputs. 

• Amazon DynamoDB: NoSQL store for workflow metadata, session states, and audit logs. 

• Amazon API Gateway: Exposing REST/HTTP APIs for external applications, UIs, or partner systems to interact with design agents. 

• Amazon CloudWatch: Metrics, and logs for agents, modes and application services. 

• Amazon QuickSight: For dashboarding asset generation metrics, agent activity, and creative pipeline performance. 

• Amazon ElastiCache (Redis): For short-term memory context. 

• Amazon Neptune: For long-term knowledge graph and relationship mapping.

Architecture Diagram 

Implementation Details 

The implementation followed an Agile, sprint-based delivery model executed over a 9-month period, involving iterative feature development, integration testing, and continuous stakeholder feedback from design leads, marketing teams, and platform architects. The engagement began with structured discovery sessions that included branding stakeholders, ML engineers, and DevOps architects to define key design automation workflows, asset governance requirements, and scalability objectives such as version control, visual consistency, and real-time integration with creative systems. 

Visual Generation Agent core components were containerized and deployed on Amazon EKS. These agents interacted with Stable Diffusion models accessible by Amazon Bedrock, enabling AI-powered generation of brand-aligned visuals such as logos, icons, and banners.  

During mid-phase development, a context memory layer was introduced to enhance generation accuracy, personalization, and workflow continuity: 

• Amazon ElastiCache (Redis) was integrated for real-time, session-level memory tracking recent prompts, selected styles, and active brand configurations. 

• Amazon Neptune served as the long-term knowledge graph storing relationships between users, brands, assets, and campaigns to drive intelligent, context-aware prompt construction and creative decision-making. 

Integration pipelines were established via Amazon API Gateway, allowing external systems such as application FE to interact with agent. Authentication and authorization were managed using IAM roles, with all endpoints secured within VPC-private subnets to meet enterprise-grade security requirements. 

Monitoring and observability were embedded using Amazon CloudWatch, capturing real-time metrics for agent execution, model inference latency, asset generation status, and API health.  

Timeline and Major Milestones: 

Months 1–2: 

• Set up Amazon EKS cluster with baseline IAM, VPC, and networking 

• Initialize Git repositories and infrastructure as code (Terraform, Helm) 

• Define agent workflows and system architecture  

Months 3–4: 

• Configure Amazon S3 (with versioning) and DynamoDB tables 

• Implement prompt ingestion logic and asset metadata schema 

• Develop modular workflow templates and API structure 

Month 5: 

• Containerize core agents 

• Stable Diffusion models configuration on Amazon Bedrock. 

• Set up CI/CD pipelines using CodePipeline, CodeBuild, and Amazon ECR 

 Months 6–7: 

• Implement agent orchestration and internal APIs 

• Integrate Bedrock model with agent. 

• Connect FE application via Amazon API Gateway 

 Months 8–9: 

• Configure Amazon CloudWatch for monitoring and alerting 

• Build dashboards using Amazon QuickSight 

• Finalize production deployment and perform UAT 

• Deliver documentation and environment handover 

Innovation and Best Practices 

The Agent Sketch platform was architected in alignment with the AWS Well-Architected Framework, emphasizing security, reliability, operational excellence, and cost optimization for AI-driven visual content generation. A container-native approach was adopted from the outset, deploying intelligent agents as loosely coupled microservices on Amazon EKS, enabling scalable execution, fault isolation, and independent versioning of each component in the design generation pipeline. 

All infrastructure resources—including networking, IAM, compute, and storage—were provisioned using Infrastructure as Code (IaC) with Terraform and Helm, ensuring reproducibility, auditability, and environment parity across development, staging, and production. This allowed teams to perform rapid rollbacks, environment cloning, and multi-region setup with minimal manual intervention. 

A key innovation in Agent Sketch was the implementation of agent-driven creative orchestration, where each functional stage of visual asset creation was abstracted into a dedicated agent. This modular architecture enabled organizations to extend or customize workflows without modifying the core engine, thereby supporting varied design operations such as logo generation, social media banners, and UI components within a single, unified platform. 

To maintain compliance and traceability, every asset generated was stored in Amazon S3 with versioning enabled, along with its prompt metadata, generation timestamp, and agent context. This provided a complete lineage of visual content, crucial for regulated industries or campaigns requiring reproducibility. Real-time logs and metrics were streamed to Amazon CloudWatch, and generation throughput and agent activity were visualized through Amazon QuickSight dashboards, enabling proactive system tuning and performance analysis. 

Collectively, these innovations and best practices positioned Agent Sketch as more than just a generative design tool. It became a highly governable, scalable, and extensible creative automation platform, capable of adapting to evolving brand requirements, multi-team workflows, and enterprise-grade operational needs. 

Results and Benefits 

Business Outcomes and Success Metrics 

• 60% reduction in manual design workload through automation of routine asset creation 

• 4x faster turnaround time for visual content generation across digital campaigns 

• 95% consistency in brand guideline adherence via automated compliance validation 

Technical Benefits 

• Containerized microservices deployed on Amazon EKS enabled modular scalability and low operational overhead 

• Stable Diffusion-based models accessed via Amazon Bedrock provide reliable inference at scale. 

• Versioned asset storage on Amazon S3 ensured traceability and rollback across the asset lifecycle 

• Real-time monitoring via Amazon CloudWatch ensured proactive issue resolution and performance tuning 

Customer Testimonial 

“Agent Sketch has completely transformed how we approach creative workflows. What used to take days of design iterations now takes minutes. The consistency it delivers across our brand assets is outstanding. The automation doesn’t replace our creatives it empowers them.” 

— UI/UX Designer 

Lessons Learned 

Challenges Overcome 

• Workflow complexity increased with multi-agent orchestration. 

• Controlling output quality through effective prompting was a key challenge when generating brand-aligned visuals 

• Maintaining secure and traceable storage was addressed by enabling S3 versioning and applying strict metadata tagging, ensuring compliance and full asset lineage with minimal overhead. 

Best Practices Identified 

• Use modular, agent-based design patterns to isolate logic and allow flexible workflow customization 

• Leverage Amazon Bedrock to access foundation models while managing versioning and safe experimentation externally. 

• Enable S3 versioning and metadata tagging to support full asset lineage, compliance, and rollback 

• Implement API Gateway + IAM for secure, controlled access to asset generation services 

Future’s Plans 

• Expand support for video and motion graphics generation using transformer-based models and multimodal prompts. 

• Integrate with customer feedback loops to improve model output based on real-world usage patterns. 

• Explore multi-language prompt support to enable localization of visuals for global marketing teams.