The default for AI in 2025 is hosted. The hosted default makes sense for some workloads and is structurally wrong for others. Worth being explicit about which workloads belong on-prem and why the inertia is hard to break.
The first article of my personal coding constitution: every change starts with a tracked issue, not a commit. I had to write this rule down for myself before I'd live by it. Then I had to write it down again for the agents I work with.
Every AI team I've talked to has the same problem: private information moves through their tools, and nobody on the team owns it. The gap is real, it's getting wider, and something will force the issue. Here's the breakdown, and who should own it.
GitHub Universe ships Agent HQ. Big-tech earnings push 2025 AI capex past a quarter trillion. Amazon cuts 14k jobs the same week. OpenAI completes its PBC restructuring. Cursor ships its own model. The week that made the AI capex-vs-jobs trade explicit.
Latency, token throughput, cost-per-request, queue depth, drift. Prometheus for the metrics, Loki for the prompts and responses with PII discipline, Tempo for the tracing across agent calls. The dashboard JSON itself as a Decisions as Code surface, projected per-environment.
The keynote pitch for agent orchestration is engineering-shaped, pipelines, queues, message buses. The framing that holds up better in production is workforce-shaped, roles, teams, supervisors, performance review. Worth being explicit about why.
What happens when every tool the agent uses is exposed via MCP, every memory access is via MCP, every observability hook is via MCP. The architecture gets cleaner, the lock-in shrinks, and a few specific design decisions become obvious.
OpenAI ships Atlas, an AI-native browser that watches everything you read. Anthropic locks in a million Google TPUs. 850 luminaries call for a superintelligence ban. The week the centralization story got load-bearing at both ends of the stack.
The headline AI pattern of 2025 is agentic-everything. The pattern that's quietly doing more useful work for actual production systems is distillation, taking a big expensive model's behavior and getting most of it from a much smaller cheaper one.
Tech debt was the unfunded liability of an engineering team for two decades. Context drift, the slow erosion of what the AI tools think you're doing, is the equivalent for AI-augmented teams. Same shape, different substrate.
Anthropic ships Haiku 4.5, then drops Skills 24 hours later, two releases pulling on the distributed-AI argument. OpenAI-Broadcom go 10 GW custom silicon. NotebookLM gets real chat. The AI-layoff drumbeat keeps building. My take.
Deployment frequency, lead time, change-failure rate, MTTR. The DORA framework worked for a decade because it measured the right things. AI-augmented engineering is bending those metrics in interesting ways and exposing the next thing to measure.
