An honest retrospective. The mistakes I made with AI in 2025-26, over-trusting model output, under-investing in evals, picking the wrong abstractions, skipping verification. The lessons I actually learned, written down so I don't have to learn them again.
Google Cloud Next launches the Gemini Enterprise Agent Platform. GPT-5.5 ships. Google commits $40B to Anthropic on Friday. DeepSeek drops a 1.6T MIT-licensed V4 the same day. April Challenger lands ugly. The frontier-lab market is now hyperscaler-vs-hyperscaler.
A semantic search is too helpful by default. It will find the personal note that shares a name with the work doc. Scope the index.
Persona isolation that stops at the application layer isn't isolation. The DB has to know the persona too, row-level filters, audit logs, the whole thing.
I spent a few weekends pushing a local-first assistant onto core-01, my M4 Max Mac Studio with 64GB of unified memory. The path that works is Apple MLX. Here is why I picked it over llama.cpp, Ollama, and vLLM, and what the tooling actually feels like in 2026.
Policies vary per tenant. Code does not. The architectural payoff of tenant-aware policy bundles. OPA-style, that don't require a tenant-aware codebase. Bundle resolution, defaults, overrides, and the audit story that makes 'what happened for tenant X' answerable without grep.
Long-term memory has to be persona-scoped from day one. Otherwise the AI brings up the wrong thing at the worst possible moment, and you stop trusting it.
The practical version of the second-brain piece, what I actually run on Apple Silicon, how notes get captured, how search and RAG work against my personal corpus, which models do which jobs, and what the daily loop feels like once it's in place.
Anthropic ships Claude Opus 4.7 with a tokenizer change that quietly raises the bill ~15-25% at the same headline price. OpenAI fires back with GPT-5.4-Cyber. MiniMax open-sources a self-evolving agent model that ends the hosted-frontier argument. Q1 layoff data lands ugly.
The tool-routing layer between an AI and the rest of your life has to know which persona is asking. Otherwise the wrong tool always shows up.
When the persona switches, four things move at the same time: the prompt, the tool surface, the memory, and the identity doing the signing. All four. Atomically.
Sixteen years in: vCAC, vRA, AWS, OpenStack, Kubernetes, Terraform, now AI agents. The honest read on what persists across the eras, idempotency, declarative state, policy-as-code, observability, and what's genuinely different about agents.
