Post

Back from speaking at @aidotengineer.bsky.social in SF last week and #googleioconnect Berlin just before. Lots of energy and cool (AI) stuff * The self-improving codebase is real * The agentic org chart * Agentic web standards * Agent sandboxing * Open models, training, fine-tuning #AIEngineer

The self-improving codebase, aka Dark Factory, is real, but old rules still apply. Good CI/CD, e2e tests & canaries directly speed up the agentic loop — skip them and humans stay the bottleneck. ie. Uber already write, review, test & merge PRs with barely any human touch

The agentic org chart. Forget managing multiple tabs — the new model is for one human to interact with one "AI manager" that manages multiple agents. Cloud dev environments let those sessions run nonstop at scale, no need to keep your laptop open.

The agentic web is getting standardized. WebMCP (Chrome): sites expose in-browser tools directly to agents, no DOM scraping. MCP Apps: servers push interactive UI back into agent chats. ARD/ai-catalog.json: agents discover tools & agents at runtime. auth.md agents register users without signup forms

Agent sandboxing is the new infra fight: Gemini Enterprise Agent Sandbox, AWS Lambda Sandboxes, k8s agent-sandbox and substrate, Nvidia OpenShell, Modal, E2B,... Workload isolation + fast snapshot/restore

Models, Training and datacenters: plenty of companies selling inference, model routing, training, post-training, fine-tuning,... on open models