Latest posts
Measured, reproducible notes on local AI agents — every claim with code you can run. Proof, not vibes.
- Local AI
Simulation First, Then Go Live: I Built a Bench That Proves an AI Harness Before It Touches Real Data
Never let an AI harness meet real data until its declared invariants have been verified — reproducibly, in simulation. I built Harness Lab to make that discipline mechanical: prove the harness in sim, then go live by rebinding one tool. The method, the five rules the bench enforces, and the proof — my agent's email path run both ways, two traces the same shape. Updated 2026-07-15: the bench is now open source, and the proof is one command.
- Local AI
The Spine, Drawn: I Made My Agent's Governance a Graph You Can Sweep — and Can't Bypass
Last post I argued the harness is a deterministic spine and the guards, not the model-judges, hold the line. This one earns the next claim: I rebuilt that spine as an explicit graph on a bench, swept every governance knob, and watched the safety line appear as a number. When governance is topology instead of an aspect, the bypass isn't a test you keep passing — it's an edge that doesn't exist.
- Local AI
The Harness Has a Spine: Why Deterministic-First Beats Pure Agentic for Local AI
Two experiment campaigns on a local, privacy-first agent — a tuning-lever sweep and a judges-on red-team. Both times, reliability and safety came from the deterministic scaffolding, not the model or the knobs around it. For local agents, the leverage is a harness with a spine.
- Local AI
Local AI Agents in 2026: The Model Isn't the Bottleneck, the Harness Is
An eagle-eye view from three experiments on local open-weight models, run on an Apple-Silicon laptop with MLX. The model on your machine is already good enough for most real work; what decides whether it works is the engineering around it — and every failure I found was a locatable, fixable harness problem, not a model limit.
- Local AI
The Buffer-Size Cliff: The One Setting That Stops Your Local AI Agent Hallucinating
A sharp quality cliff at buffer_size ≈ 0.5 × prompt_tokens makes local AI agents stop calling their tools and start hallucinating — measured on six open-weight instruct models on Apple Silicon.