AI agent
An AI agent.
Notes
What makes a good harness? (Notes from February 1, 2026)
- Elements such as static type checking and LSP that can be beneficially connected to an LLM seem to make useful, sustainable harnesses.
- By contrast, the advantages of Vector database-based RAG over Agentic search are rapidly fading in many situations, because it offers little advantage in terms of “beneficially connecting elements.” Of course, embeddings remain useful when there is enough data.
Not building an agent (notes from around October 2025):
- Perhaps a sufficiently widely used general-purpose agent will emerge during 2026. Just as web browsers became widespread in the 1990s, might general-purpose agents become popular in 2026? If so, agents operating within individual websites or apps will find themselves in a strange position.
- In the late 1990s, several small companies were making web browsers. They should have been making good websites instead.
Articles
- 2024-12-20 - Building effective agents \ Anthropic