Series 2 of 4 · 8 episodes · 1 season

HARNESS
ENGINEERING

The model is 10%.
The harness is 90%.
Design the system around it.

Start with Episode 01 · Why Your Agent Fails → See the full map ↓
Runtime, not prompt 0.95⁷ = 70% compound reliability
Evidence · Selected

Three receipts. One season.

01
Not the model
Every post-mortem.

The hallucination wasn't a hallucination. The window was stale. The tools were wrong. The structure was missing. That's harness.

Field notes · 2024–2026
02
0.95⁷ = 70%
Compound reliability.

A 95%-good step chained seven times succeeds 70% of the time. The harness is what stops the math from killing you.

Math, not opinion.
03
31 → 9
Fewer moving parts. Half the errors. One week.

A team spent three months hunting a smarter model. It didn't help. Then a new hire made three small fixes to the plumbing around the model — nothing to the model itself. Errors dropped by half in five days. Same model, same team. The leverage was never in the model. It was in what they built around it.

E03 · Three Small Changes
SEASON 018 episodes

Act 01What the Harness Owns

By the end, you can sketch your own harness on a napkin, hold the five paradoxes under pressure, price the program for a skeptical CFO, run the nine-day audit, and rebuild the org around it — and tell scaffolding apart from the forever work.

The MHTE stack · the harness is the moat
01Modelrented, commodifying
02Toolsowned by IT
03Environmentowned by data
04 · yoursHarnessorchestration · retries · guardrails

Three layers already have owners. The harness has none. That's the opening.

T01 Why Your Agent Fails 2 AM. The most capable model money buys just declared itself done at invoice 847 of 2,347. The autopsy says model. It isn't. If you only read one — start here Ep01 → 4 signatures READ T02 Inside the Harness You bought a harness. Nobody in the room can name what's inside it. Four layers, five clusters, one runtime — and why guardrails belong in IAM, not the system prompt. Ep02 → 4 layers READ T03 Three Small Changes A quarter chasing a better model. Then three fixes to the plumbing around it. Errors dropped by half in five days, same model. Ep03 → 3 patterns READ T04 Five Paradoxes You Have to Hold Smarter models don't behave steadier. More guardrails buy more freedom. Five inversions that separate teams that ship from teams that keep piloting. Ep04 → 5 paradoxes READ T05 What It Costs. What It Returns. Month four. The CFO opens the spend dashboard before the meeting starts. You have ninety seconds. Five cost centers, the reliability dividend, three signs you've crossed break-even. CFO-ready Ep05 → 5 cost centers READ T06 The Monday Morning Kit Most harness advice lives at 30,000 feet. This one lives at your keyboard. A 9-day audit, a 5-rung ladder, four tickets to file this sprint — with the exact wording. Ep06 → 9-day audit READ T07 The Org That Owns the Harness Three layers already have owners. The harness has none. Five companies rebuilt around that fact — and out-shipped much larger teams. Ep07 → 5 companies READ T08 When the Harness Becomes the Habit Every pattern here has a shelf life. That isn't the bug — it's the engine. Five capabilities the model absorbed; five it never will. Season finale Ep08 → 5+5 patterns READ

The harness is the contract you make with the model. And the only thing you actually ship.

Start with Topic 01 → See the full map ↓
The syllabus. · Three next-hops — pick your angle.
Agentic Stack · Ep 01 Prompt vs. Context The iceberg. The user's message is less than 1% of what the model actually sees. Read → AI Evals · Ep 01 Benchmarks ≠ Evals The needle. Benchmarks measure the model. Evals measure what you shipped. Read → AI PM OS · Ep 01 Why AI PM ≠ SaaS PM The compass. The job changed underneath the title. Here's the new one. Read →