University of Edinburgh Prototypes AI Agent System in EdWeb

• University of Edinburgh Prototypes AI Agent System in EdWeb (Experimental)
• Source: Integrating ELM with EdWeb – Building an AI tool for publishers by Mostafa Ebid (22 August 2025)

TL;DR: During a summer internship, the University of Edinburgh developed a prototype AI assistant inside its Drupal-based CMS, EdWeb. Built by Mostafa Ebid, the project integrates the University’s in-house language models (ELM) with the Drupal AI and AI Agents modules. The system uses specialized agents—such as a style guide agent and site maintenance agent—drawing on trusted University resources like its editorial style guide and acronym database.

User testing found publishers valued clearer headings, better link text, and layout suggestions, though they preferred adapting AI feedback rather than using it directly. Future iterations will refine orchestration patterns, improve semantic search, and enhance transparency through clearer source attribution. Detailed story below:

The University of Edinburgh has unveiled a prototype AI tool built directly into its Drupal-powered content management system, EdWeb, designed to help web publishers improve content writing, page structure and compliance with editorial standards.

The project, developed during an internship by Mostafa Ebid, integrates the University's internal language models (ELM) with EdWeb through the Drupal AI module and a new "AI Agents" framework. The tool aims to provide trustworthy, context-aware assistance for publishers by drawing exclusively on University resources, including its editorial style guide, acronym database and content design best practices.

Unlike general AI writing tools, which often hallucinate, alter tone or give shallow feedback, the Edinburgh prototype uses a multi-agent orchestration system. Specialized agents-such as a style guide agent, page structure agent and site maintenance agent-access tailored tools like semantic search (RAG), sitemap analysis and JSON-based style guide lookups to deliver precise, publisher-centered responses.

User testing revealed publishers valued the AI for suggesting clearer headings, better link text and alternative page layouts, though they preferred adapting its advice rather than using it verbatim. Feedback also highlighted a need for clearer source attribution, prompting future iterations to display references alongside AI-generated suggestions.

Ebid initially experimented with a project-manager orchestration pattern but scaled back to a leaner orchestrator-to-agent design after finding "overplanning" diluted responses. Future work will explore alternative orchestration strategies, optimize system prompts using techniques like ReAct to reduce hallucination, and improve semantic search for detecting duplicate content.

"This is just version two," Ebid wrote. "The prototype can be refined through new tools, better orchestration, and improved system prompts to ensure it truly helps publishers in their daily work."

The effort underscores how Drupal-based CMS platforms like EdWeb can embed AI assistance directly where content creators work, reducing reliance on external tools and tailoring feedback to institutional standards.