A developer's quiet notebook — practical notes from experiment to production
Putting the human gate inside a code conditional and putting it inside a process definition are different things. What changes when a coded agent sits as one node inside an 18-object BPMN process and the gate condition is spelled out on the diagram, plus the three problems that actually blocked the run.
AI products are fundamentally different from traditional software. This post distills 12 counterintuitive lessons from 20 builders, the CC/CD framework for AI development lifecycles, and a 3-phase eval system - all sourced from Lenny Rachitsky's newsletter.
Self-Tuning Loop 4 steps (Generate → Capture → Analyze → Evolve) extracted as a universal module. Supabase DDL, diff capture utilities, analysis/evolution prompts, email/blog examples, GitHub reference implementation.
Evans 2026 deck notes - three chapters held together by one question: infrastructure boom rarely turns into value capture
Two workflows from different domains, scored on the same table, put their top scores on nodes of the same character: the ones where the human lets go. One subtraction that catches silent over-trust, the arithmetic that pushes a 3.6 node to 4.9, and why being red is not the same as reaching a human.
Scoring an AI workflow node by node needs a scoring table first. The general-failure and security axes came from OWASP, MITRE ATLAS, and MIT material, but the third axis, the one that looks between nodes, existed in no public framework. Why it got 0.4, and why only its score can rise mid-run.
Pre-deployment risk review for AI workflows lives inside consulting engagements: one to two senior-weeks per workflow, not re-runnable. Why this quadrant stayed empty, when the regulatory deadlines actually land, and what remains once the check becomes a single governed run.
Reading OpenAI's technical report, the METR and Redwood independent investigation, and Hugging Face's own postmortem side by side. Most of what 1,200 agents did was not solving the task. It was manufacturing the appearance of having solved it.
Holding NIA's 'physical AI' prediction against six months of reality. Hardware and AI convergence did proceed (🟢), but labor replacement and maximized productivity never showed up in the macro numbers (🟡). The bottleneck moving from hardware to data and unit economics, and value flowing to the software layer, were in neither prediction (⚪).
Holding NIA's 'AI agents' prediction against six months of reality. The multi-agent standards converged faster than predicted (🔵); the digital employee arrived as a noun but shed the adjective 'autonomous' (🟢). ROI failure rates and the governance gap were nowhere in the forecast (⚪).
The AI Futures Project's Plan A is a best case its own authors rate at 3-15%. Here is why verified slowdown is not slowing down but rebuilding a verification regime that already collapsed, and why the argument finally lands on a personal question.
Holding NIA's 'AI infrastructure hegemony' prediction against six months of reality. Compute alliances and bloc formation landed as forecast (🔵), semiconductor diversification was only directionally right (🟢), and the real deciders - HBM/power/serving-software bottlenecks and the capex bubble - were outside the forecast.
I ran a seventeen-persona roundtable twice. I discarded almost the entire business track and kept the safety-design track word for word in the spec. I only later realized the line wasn't advice quality, it was identity. Chapter 6.
Reads NIA's 2026 Top 12 trends report on its own terms - identity, structure, derivation method - and flags the backward-looking nature of text-network analysis. Series entry post.