Field notes on agents, RAG, LangGraph, and the unglamorous work of shipping AI that holds up under real users. Honest, code-first, and short on hype.
// topicstap to filter
// posts2 published
Long chats eventually overflow the model's context window. Using a small fitness-coach agent, this walks through keeping LangGraph in-bounds with trim_messages — the strategies, the token budget, and the tool-call boundary gotchas most tutorials skip.
Most of an agent's latency isn't thinking — it's asking the model to make decisions a plain if-statement could make. Here's how we moved routing out of the LLM and halved runtime without dumbing anything down.
// in the pipeline3 drafts
A pattern playbook for recovery, retries, and graceful degradation in tool-using agents.
Why naive RAG fails on multi-entity questions and how to fix it without exploding latency.
A small, brutally honest test rig that catches regressions before users do.