AI Development and Agents
The person has been developing and deploying AI agents on Cloudflare Workers and Durable Objects, building an MCP server that exposes both read tools (search_people, get_person, get_page, ask_entity) and owner-only write tools.[1] They noted that agents are difficult to introduce to non-engineers unless they are running in the cloud.[2]
A key architectural challenge arose from the mismatch between their core engine's synchronous Store interface (used by grounded Q&A and citation validation) and D1's asynchronous-only database operations, which prevented direct use of D1 as a data source.[1] Additionally, they encountered a pitfall where the D1 seed export included the meta table, causing a fresh database to have an index_fingerprint that matched seeded chunks; the first composition workflow then skipped reindexing, leaving the vector index permanently stale.[3]
The person has experimented with collective intelligence by using LLMs to create multiple agent personas, enabling a simulated form of collective intelligence.[4] They cited an example where prediction markets like Polymarket yielded higher accuracy than a handful of financial analysts, drawing a parallel to the multi-agent approach.[4]
For multimodal search, they use Gemini Embedding 2 to store vectors for text, images, video, and audio in a database, enabling semantic retrieval across formats with a single query (e.g., finding all content about "cats").[5] They recommended a specific prompt for Claude Code to set up such embedding storage and a corresponding search skill.[5]
They observed that AI-generated artifacts (such as UI mockups) have become used unconsciously in their workflow.[6] Despite the engagement and enjoyment of tinkering with AI, they also acknowledged a recurring feeling of "so what" that intrudes on the excitement, reflecting a persistent existential questioning about the purpose of the work.[7]