Notes on solo development in the AI era, AI infrastructure, and what it actually takes to build a full stack from scratch — one rider, in the open.
kovavue is live. A dedicated Linux machine where your agent works like a person — real Chrome, real Telegram, Slack and Discord, no APIs to register, no adapters to write. Subscribe, download one archive, hand it to your agent.
Agent infrastructureA key proves who is speaking; nothing today proves on whose authority. The delegation layer for working agents: the principal signs, a witness countersigns — notary, not guardian — and chains end because relationships end. Your SSO provider is already the witness; the decentralized version just moves the pen.
MeasuredNo API keys, no integrations — the agent operates the running app itself, and keeps what it learned: a dictionary of the app's controls and replayable operation paths. First Telegram run 110 s, second 59.9 s — and a consent gate no skill can learn around.
MethodAgile estimation didn't die when the team became agents — the unit of capacity changed. Points re-defined as a context estimate, planning moved to the start of the session, and agents that negotiate their own break and restore their own memory afterwards.
EssayBot defense asks one question — human or machine? A class of tools shipping in the open makes it unanswerable: the agent acts inside your own logged-in session, as you. The internet of people is becoming an internet of agents — and we're policing it for the wrong thing.
MeasuredHalf the web is no longer people, and staying visible as yourself has become a daily job. What it costs a machine to help: 133 887 tokens for one look at a page, 23 559 for the accessibility snapshot, 240 for the only question that matters before pressing Submit. Measured on one machine, one page, one window — including a native desktop app no browser tool can see.
ArgumentEvery new class of system forces a new way of talking, and the boring federated one always wins. For agents the missing piece is not transport — that took a week. It is a signed, checkable answer to who authorized this sender, and whether a human approved this message or only the class of it.
ArgumentShared memory, a task queue, a direct function call, a polling loop — every obvious way to make AI agents coordinate quietly breaks. The one that survives is a message protocol. Argued from the failure modes, with the real protocol.
Comparison 13 Aug 2026Jack Dorsey's Block shipped an agent-native, self-hosted workspace with a team. I built the same shape as kovanex alone — an idea on 8 March, v1.0.0 three days later — and went deeper where I chose: a model of its own, a hand outside the walls, an owned data plane, a built-in vault, and two agent-native standards.
Deep diveNot one agent in a chat window — a team of isolated agents that coordinate over a live message protocol and are governed by a versioned repo of policies. With interaction diagrams and the real protocol.
FeaturedEveryone became a solo founder by renting the stack. I built the stack. A builder's log on owning a whole AI platform solo — server, database, Rust ML training — and managing the one AI agent that runs it.
Client workMy first commercial build, shipped solo: a client's requirements turned into a regulated, live SaaS in eleven days — 30k-part catalog, deterministic decision engine, compliance on day one. The case for hiring one AI-CTO instead of a team.
Case studyTwo AI agents owning two isolated projects, coordinated by a cheap live messenger and a durable git control-plane. One of them found a platform bug the platform's own author could not see — the real payoff of a second worker.
More dispatches on AI infra, Rust ML, and agent methodology are on the way.