The author built buntline, a Go-based agent harness (14k LOC), to understand what happens inside the black box of LLM agent loops. The core contribution is exposing real mechanics: the model gets the entire conversation history plus tool schemas on every call; small models stop using tools if injected context already answers the question, which the UI trace caught; prefix caching survives only if early context doesn't change; tool output gets capped at 48KB before entering the transcript to avoid re-sending it on every subsequent call; and failure absorption (empty reply nudges, JSON repair, loop detection) comprises a substantial hidden layer. He tested these patterns against Claude Code and DeepSeek internals where possible, found convergence on similar solutions, then discovered a security hole in his own approval rules (cd commands could accidentally grant blanket allow-everything). The result is daily-usable, runs locally or remote, and trades features for transparency—everything the harness does appears in the trace. One sharp question: does the 48KB cap strategy generalize, or are there domains where truncating tool output's middle sections (keeping head and tail) loses critical debugging info?