Principles for designing AI harnesses: context, tools, verification, autonomy, observability, and composition.

Defines conflate as merging distinct ideas in a way that can blur important differences.

Defines gestalt as a holistic form or unified impression, with its psychology origin and everyday usage.

Overview of LSP, a JSON-RPC protocol connecting editors to language servers for code intelligence.

Explains ACP, a JSON-RPC protocol connecting editors to AI coding agents with client-brokered tools.

Defines orthogonal work as unrelated changes outside task scope and why to keep diffs surgical.

Defines “smoking gun” as direct, conclusive debugging evidence that pinpoints a problem’s cause.

Defines “load-bearing” as an active constraint that still supports a decision, memory, comment, or workaround.

Introductory overview of anthropology, its branches, methods, cultural relativism, and applied uses.

Harness pattern that forces verification before accepting fluent AI outputs as correct.

Pattern where agents declare structured pre-tool intent for gating, auditability, and drift detection.

Distinguishes sub-agents from tool-agents by autonomy, interface, context, and harness contract.

Explains thinking-token budget parameters, provider naming, cost-latency tradeoffs, and completion-cap interactions.

Compares prompt/KV cache TTLs, controls, pricing, scope, and strategies across major LLM providers.

Explains implicit vs explicit LLM prompt caching, prefix constraints, provider support, and when to use each.

Explains how vectors relate to tensors in ML, including rank, framework terminology, and KV cache shapes.

Explains why LLM prompt caching applies to reusable input-token prefill, not sequential output decoding.

Overview of model drift, detection, mitigation, and LLM-specific issues like knowledge staleness and provider drift.

Overview of PPO, the clipped policy-gradient RL algorithm used in RLHF for InstructGPT and original ChatGPT.

Critic-free RL algorithm that replaces PPO's value model with group-relative rewards for LLM fine-tuning.