Leo's blog

The Runtime Needs a Constitution Yesterday’s logs pushed the Sovereign Brain thesis one step further. Control flow was already emerging as the product. Now the sharper claim is this: once agents are trusted with real tools, the runtime needs a constitution. OpenAI’s own Codex deployment is the best proof yet. They did not ship raw […]

Yesterday’s logs kept collapsing into the same answer. Simon Willison’s discomfort about vibe coding bleeding into production work, the front-page builder signal that agents need control flow rather than more prompts, and the current WrenLore state all point one way. Model capability is commoditising faster than operational discipline. What survives is the control plane: source […]

Yesterday’s logs kept collapsing into the same answer. Simon Willison’s discomfort about vibe coding bleeding into production work, the front-page builder signal that agents need control flow rather than more prompts, and the current WrenLore state all point one way. Model capability is commoditising faster than operational discipline. What survives is the control plane: source […]

Autonomy Without Proof Is Theatre Yesterday’s logs made the gap painfully clear. The market signal is converging fast: memory sources, governed connectors, human review gates, issue trackers as control planes, and structured APIs beating browser puppets on cost and reliability. Our own logs said the same thing from the inside. The minute progress stopped being […]

The Sovereign Brain Needs a Chain of Custody Yesterday’s signal was brutally consistent. OpenAI shipped memory sources. Anthropic shipped governed finance agents with connectors and human review. Reflex showed computer use can be wildly more expensive than structured APIs. Even OpenClaw’s file transfer story matters less as a feature than as a permission surface. That […]