Essays by Mark Ferraz on AI infrastructure, agents, memory, and the gap between demos and production.
AI capability does not produce delegation. Trust does, and trust is a function of four variables: continuity, shared context, boundary clarity, and an inspectable track record. The industry races capability, the one variable that does not move the relationship. Why the user pays for the freedom to stop looking, and why every variable that earns delegation is a layer problem, not a model problem.
AI coding agents collapsed the cost of generating software while the cost of securing it climbed. Frontier models with real cybersecurity capability are gated behind access controls and metered pricing, which binds legitimate defenders while sophisticated attackers already have the tools. The unit economics of defense in the age of AI, and why concentrating defensive capability leaves critical…
Frontier intelligence in regulated work is insurance against catastrophic error, not labor arbitrage. When a model can be withdrawn by a party outside your contract and beyond your auditor's reach, that dependency stops being a risk to price and becomes a control finding to remediate. The architecture that quarantines frontier reliance to the few high-consequence decisions that justify it, and…
Generation got cheap. Judgment did not. Why discernment, taste, and eval discipline are the assets that compound in the next decade of applied AI, and why agent slop ends careers faster than any model release.
A decade of building systems inside the energy majors (Chevron, Marathon, Southwestern Energy): what AI changes in those environments, what it doesn't, and what the operator-architect carries that frameworks cannot.
Why model-level tuning isn't enough: alignment lands at the user, not in the lab. The common grounding problem. Why the labs aren't building the depth layer, and why the people who are tend to be the ones the platforms overlook. What a real harness around the model has to hold for delegation to be trusted.
Bitemporal claims that supersede instead of overwrite. Access boundaries that travel with the data. Why most enterprise AI memory leaks at the API edge, and what to build instead.
Why public training data converges to the median, what that means for solo operators using AI for research, and why human judgment must precede generation.
A 25-year practitioner's view of the divergence forming between operators who use AI to extend judgment versus those who use it to substitute for it.
The semantic inflation of the word 'agent' is masking real engineering tradeoffs. What a production agent actually requires versus what most vendors are shipping.
Building production software with autonomous coding agents instead of a consistent team. What works, what breaks, and the operator discipline required.
The distinction between conversational memory, vector retrieval, and structured knowledge graphs, and why most LLM applications are quietly building the wrong thing.
Regulated-industry requirements for AI agents (auditability, policy enforcement, deterministic guardrails), and why most commercial offerings fall short.
Foundation models are commoditizing fast. The durable moat is the harness, orchestration, evaluation, memory, governance, not the weights themselves.