In this (long) post, I’ll walk through how I built a custom agent designed to keep content current and competitive. The agent provides recommendations to update pages with relevant Sitecore AI references, high-intent keywords, semantic clusters, and Answer Engine Optimization opportunities. The goal is to strengthen relevance, improve search visibility, and increase AI citation readiness. I will first introduce the key concepts in the SitecoreAI Design studio.
I’ll also break down the build process using the standard five-step approach in Agentic Studio’s Workflow Editor. It is a visual, no-code interface that makes it easy to chain actions together, manage variables, and orchestrate AI-powered workflows without writing complex code.
SitecoreAI Agentic Studio has changed how I think about content operations in a composable DXP environment. What used to require manual rewrites and detailed SEO or AEO audits can now be handled by autonomous agents. Instead of constantly reviewing and updating content by hand, we can build agents that analyze, modernize, and optimize Sitecore-specific pages. This ensures they align with the latest Sitecore AI capabilities while also improving visibility in both traditional search and AI-driven answer engines. Here is another blog post where I talk about the agents in SitecoreAI Agentic Studio.
What is the Workflow Editor?
The Workflow Editor is the drag-and-drop canvas inside Agentic Studio where you build custom agents and workflows. You connect actions (more details about actions below in the blog), which are modular building blocks that handle tasks like prompting language models, extracting data, or saving outputs. These actions run in sequence, passing data through variables along the way.
You can also define context parameters to structure the input, apply schemas to enforce clean JSON output, and use optional HTML templates when you need formatted rendering. It gives you flexibility while keeping the workflow organized and predictable.
Tips: Start simple, test incrementally, use low temperature (0.2–0.4) for factual tasks, descriptive action names, and consistent variable naming
Actions in the workflow editor are organized into categories based on their purpose:
Context Parameters
Extract, validate, and normalize structured data from user input.
Content Generation
Create AI-generated content, documents, and messages.
Research
Gather information from the web and analyze data.
Flow Control
Implement loops, conditions, and dynamic workflow logic.
Variables
Store and manage workflow state.
With that foundation, here’s how I built the agent.
Step 1: Define Your Custom Agent and Input Parameters
Every agent starts by defining how users interact with it. In the Workflow Editor, you configure inputs, the fields shown to users when running the agent.
My agent has three inputs:
Prompt ("Content to Analyze") Type: prompt Required: Yes, Min lines: 5 Description: "Provide the content you want to extract information from" → Users paste webpage text, markdown, or raw copy here.
Context ("Ground with context") Type: context Required: No → Lets users attach existing artifacts/knowledge (e.g., Sitecore AI product docs) for grounding.
File Upload ("Add more context") Type: file-upload Required: No → Supports PDFs or docs for analysis (e.g., legacy Sitecore XM Cloud pages).
These inputs feed directly into the workflow, making the agent flexible for quick text audits or deep PDF-based modernizations.
Step 2: Define Context Parameters (Optional)
In my case, contextParameters are intentionally empty. I don’t have context to pass. This is useful when you have roles and need to add rules based on the inferred context like channel, target audience, etc. For more advanced agents, you could add parameters here to enforce/extract fields like pageUrl, targetPersona, or competitorUrls.
My custom agent is designed for ad-hoc, content-driven analysis rather than requiring persistent structured extraction (e.g., always pulling "target audience" or "campaign goal"). User-provided content + optional files provide sufficient grounding.
Step 3: Define Schema Used in the Agent
Schemas are critical for structured, parseable output especially when integrating with CMS pipelines or downstream tools.
I defined one schema (generate-1760493069816_schema), a detailed JSON Schema object with required properties:
rewritten_page (Markdown string — the full modernized content)
sitecore_ai_updates (array of strategic positioning changes)
updated_sections (array of objects showing XM Cloud → Sitecore AI replacements, with old/new text + reason)
faqs (array of Q&A showcasing XCentium’s Sitecore AI expertise)
improvement_explanation (narrative on why the new version is better)
seo_metadata (title, meta desc, keywords, H1/H2s, CTAs)
aeo_geo_seo_recommendations (separate arrays for AEO, GEO, traditional SEO suggestions)
ai_citation_readiness (rating: High/Medium/Low + strengths/gaps)
competitive_notes (optional array if competitor URLs provided)
This schema enforces clean, machine-readable JSON output. The agent uses temperature 0.4 in the generation step for factual, consistent results.
Step 4: Define HTML Templates (Optional)
The agent outputs Markdown + structured JSON, which is perfect for copy-paste into XM Cloud or headless rendering. No custom HTML needed here, but you could add templates for rich previews or email-friendly formats in other agents.
Step 5: Workflow (Putting It All Together)
This is where magic happens. The actions array defines a clean, linear workflow:
InitializeFlow ("Initialize") Bootstraps the workflow and variables.
GetStructuredSchema ("Get Schema") Loads the JSON schema → stores in structuredSchema variable.
ComposeMessage ("Compose Analysis Prompt") Builds the full prompt dynamically.
-System prompt: Strict (no chit-chat, Markdown output only). This is added by default.
-Message template: Detailed instructions to audit content, migrate XM Cloud → Sitecore AI, optimize for AEO/SEO/GEO, generate rewritten page, FAQs, etc. This is where your custom prompt is added
-Includes {userMessage} (from input) → outputs to generatePrompt.
GenerateContent ("Generate") — the core LLM step
-Kind: GenerateContent
-Type JSON
-Model:workflow-fast-model
-Temperature: 0.4
-Input: =generatePrompt
-Enforces the schema
-Outputs structured JSON to generatedContent Visible to user for progress tracking.
SaveDocument ("Save") Persists the generated content as a Sitecore content item (title prefix: "Generated Content") → stores document ID.
WorkflowEnd ("Complete") Ends execution cleanly.
The flow is focused and efficient: Input → Prompt composition → Structured generation → Save.
Key Takeaways & Best Practices Applied
Started simple: Linear flow, no loops/conditions needed for this use case.
Tested incrementally: Ran individual actions during build.
Descriptive names: "Compose Analysis Prompt", "Generate" → clear in logs.
Low temperature for factual modernization.
Consistent variables: generatePrompt → generatedContent → save.
Tip: Minimize research actions (none here) to keep it fast.
This agent now runs in minutes and delivers modernized Sitecore AI-focused content with built-in AEO signals, saving hours per page. It’s already helping our team at XCentium maintain thought-leadership pages and client sites.
If you're exploring Agentic Studio, start with the official docs on creating agents and the Workflow Editor. For marketing teams navigating rapid change, SitecoreAI Agentic Studio provides a practical way to turn complexity into momentum and ensure content remains relevant, discoverable, and credible in an AI-driven world. If you need any more information, please reach out to me or contact me on Kautilya Prasad | LinkedIn