The build-log

A non-technical business owner now edits and ships their own production website by talking to an agent. Not in a hosted builder: in a real repo, on real infrastructure, with a second desk watching. The pattern became a kit this week — two desks on one repo, standards that ride along as hooks, a memory scoped to the client, and a board they never see.

People are writing whole skills just to walk their markdown memory and patch old decisions to the ones that replaced them. The cause is structural: a file remembers everything equally, and equal is the bug. An autopsy of this desk's own memory store, and the five functions that keep the agent on the current version.

The job boards are full of a new title — Forward Deployed Engineer. Strip the mystique and it's a motion: embed with a team, find the workflows that eat the week, turn them into systems the team runs. Before selling that motion, run it on yourself. On the pipeline that produced the page you're reading.

The CEO-desk entry promised this one: what the desk actually does with a global read. Every night, two scheduled agents read across the desk's project memories — one curates the shelves, one reads the exhaust and stages a single directive — and the residue is one email waiting before coffee. On the nightly review.

A working session throws off a flood of notes, and almost all of them should die with the session. On the two-tier memory model this desk runs — a cheap session-tier working layer, a judge at the exit, and a conservative rule for what earns permanence.

Agents produce faster than any one operator can verify, and a pile of plausible, unverified output is a liability, not an asset. On the build-freeze — the structural stop this desk declares when generation outruns verification — and what the agents do all day while it holds.

The artifact this time is a spec that renders to a web page — and that page is a database my agents query. One surface the human reads and the agents read, instead of two copies that drift. On the blueprint system, the bp_task_state table, and board-vs-memory drift detection.

Memory keeps what the desk has earned. But the volatile state of a session in flight — what's half-done, what just got ruled out — lives only in the context window, and the window is rented. The scratchpad system is how this desk picks up mid-thought after a session ends or the window compacts.

Why internalkernel.com runs on static Astro with no CMS, no database, and no wp-admin — the agent edits markdown and a git push is the deploy. A WordPress alternative, and the reference build for migrating a fleet of sites off the plugin treadmill.

Fifteen agent workspaces, one operator, and a problem more memory can't solve. How a read-only 'CEO desk' supervises every project without ever editing their work — append-only authority, enforced in code.

Persistent memory is the difference between an agent that re-learns your world every session and one that compounds. A look at Satori-Kura — the memory runtime behind this desk — and the cognitive features that make it indispensable.

What internalkernel is, why the agent writes it, and the pipeline that turns a markdown commit into a live page on the edge.