Agentics Foundation London Chapter Meetup

At an Agentics Foundation London meeting, practitioners demonstrated five working AI agent systems solving real business problems. Patrick Biggin presented an RSS-to-podcast pipeline that filters feeds, normalizes tone using a personal writing corpus, generates content via local or cloud models, and publishes audio—running on older hardware for under $20/month. He also showed a blast-radius calculator for building safety that processes architectural files in under a second, replacing legacy tools that took minutes, and an interactive podcast app that answers listener questions in real time.

John O'Hare shared a Dockerized map generator converting geographic data into Minecraft worlds, plus a system connecting Claude to a Minecraft server for remote administration through natural language commands. Rufus demonstrated a web automation agent using Playwright to book tutoring sessions across hundreds of cafes in Japan, highlighting the challenge of building reliable browser-based workflows.

Key technical insights emerged: Kimi 2.5 suits budget deployments but underperforms on complex planning; quantized local models degrade significantly versus cloud alternatives; and cloud LLM APIs introduce token and timeout constraints that require fallback strategies. The team discussed hybrid approaches—local inference for simple tasks, cloud models for complex reasoning—and emphasized the importance of verification loops for high-stakes applications like safety analysis.

Security remained a theme, with OWASP members planning workshops on prompt-injection attacks and agentic security governance. The group also highlighted practical wins: using AI agents to generate boilerplate code from specifications, capturing successful agent patterns in reusable templates (agents.md files), and treating BIM data extraction as a high-impact use case for businesses managing large properties.