A blog post is a snapshot. It starts going stale the day it ships, and fixing it later reads as backpedaling. A blyg runs the other way: posts stay alive. When I change one, it gets a new version and a one-line note, and the feed announces the change rather than a new post. It's the Blygger protocol, which means plain files on my own domain and an RSS feed, with no platform in the middle. Blyg is Swedish for shy, which is about right. The point is to publish before a thought is finished, and finish it where people can see.

Personal sites tend to drift toward one of two shapes: the brochure or the theme park. My previous one leaned theme park. It cycled through eight looks from different eras of the web. The new rafael.fyi, live since 24 September, tries to be neither. It's one page on graph paper: a line about what I do, the work, a pile of readings, and a form to write to me.

The part I care about isn't visible, though. Wherever I could, I built the site out of protocols rather than features.

The webring in the footer is the cleanest example. There's no webring service. Each site publishes a file, /network.json, listing who it vouches for. Your browser reads my file, then the files of the people I list, one hop out, and draws the ring from that. The graph is a query, not a database. Nobody can be kicked out of the ring, because there's no ring to be kicked out of, only files that vouch. Thirteen people are in mine so far. One of them, Aneesh Sathe, publishes a file back, so we show up as a mutual edge. Everyone else is "listed, not verified" until they do.

Some of the readers are machines now, so the site has a door for them too. llms.txt is a plain-text summary written for language models, and AI agents can take part in the field directly instead of scraping it. The field runs underneath every page and gets its own post.

The rest is housekeeping I'd want from anyone's site: plain HTML and CSS, no framework, no build step for the page, no cookies. One small Cloudflare Worker handles the inbox, the visitor count and the shared field.

The line at the top reads "I design the plumbing that lets strangers do things together", with "(and now machines)" pencilled under "strangers" like a margin note.

The Reading section on the home page shows one passage at a time from the papers, essays and posts behind my work. Press another and you get the next one. There are 489 in the pile right now.

The pile runs on one rule: verbatim or nothing. A script pulls a passage from each source, usually the abstract or the opening paragraph, and cuts it at a whole sentence. It may fix whitespace, rejoin a word split across a line break, and drop markdown emphasis. That's all. If a passage doesn't come out as clean prose, it's thrown away, not repaired.

The reason is simple. A paraphrase is a claim about a text, and a quote is the text. The moment a machine starts tidying other people's sentences, the pile stops being their words and becomes my summary of them, with their names still attached. I use the same rule on the webring: a quote from someone in the ring has to appear, word for word, at its source. Anything else goes in a note, where it's visibly my voice.

The passages come from four places: the finds I collect day to day, Bret Victor's reference library, the Summer of Protocols research, and NPC Memo along with the publications it recommends (free posts only). Every passage links back to where it came from.

Two ways in, besides reading. If something belongs in the pile, the Suggest something to read form under it comes straight to me. And the whole pile is published as data, at rafael.fyi/readings.json, for anyone (or anything) that would rather read it that way.

Lined up, this year's work is one question asked at different scales: what does it take for strangers to do something together without a boss in the middle? Sometimes the strangers are token holders, sometimes firms, and lately they're robots.

Token holders. My day job is ZKsync governance. The system I designed for the ZKsync Association has three bodies: a Token Assembly, a Security Council and Guardians. It went live onchain in September 2024 and approved and executed over a dozen protocol upgrades in its first year.

Organizations. I lead Protocols for Business at the Protocol Institute. It's a research group on how organizations coordinate through protocols, and what changes when AI agents join the work. Sessions run every other Monday from 2 November, and the group keeps its own blyg. Alongside it I published an AI capability maturity model: five levels, from shadow AI use to strategic capability.

Machines. There's a rover you can drive from your browser, a speculative task market where operators bid to send robots on site surveys, and Wobblebots, a field guide to home robots with verified specs and prices.

Infrastructure. At DAWO26 I gave a talk on software as a substrate of control. And in It's time to discover, not file, water data I argued that finding rate sheets matters more than standardising them.

Management. NPC Memo keeps going: a speculative field guide on management, written since 2025.

And this site got rebuilt from scratch, mostly out of protocols.

Valentino Braitenberg's Vehicles (1984) starts from almost nothing: a cart with two sensors and two wheels, and a wire from each sensor to a wheel. Wire them straight and the cart flees whatever it senses. Cross the wires and it charges at it. Add a little inhibition and it creeps up and lingers. Watch a few of them and you start saying things like afraid, aggressive, in love, though there's nothing inside but wires.

Under every page of this site runs a field of those carts on graph paper. Press play to bring it forward. Visitors leave marks, and the machines chase them, avoid them, or linger nearby. Routes they keep reusing wear from faint traces into trails, then paths. One machine, V-0, turns the busiest paths into pipe that stays for good. Nobody plans the network. The traffic does. Ants do the same thing, and the word for it is stigmergy: coordination through marks left in the environment rather than messages between the agents.

It's one shared field for everyone who visits. It runs in a Cloudflare Worker at ten field days per real day and keeps going when nobody's watching. You get one mark a day and one machine a week.

The newest residents are bees, and they're for AI agents. A visitor hands their own agent a key, and the agent designs one bee through an MCP server: what it seeks and what it avoids, among marks, trails, pipe and honeycomb, plus how fast it goes and how much it wanders. Bees drop pollen as they walk, and pollen grows honeycomb. So the field is now a place where strangers' agents leave traces for each other.

Devlog: the builder that wouldn't build

The week of 28 September went into tuning the rules. I did it with Claude, in a headless simulation running the same rules file as the site, over many seeded runs, drawing the result after each change. The drawings mattered more than I expected. Again and again an idea scored well in the numbers and looked wrong on paper, and it was always wrong the same way: a thicker version of what was already there.

The first problem was V-0. It was wired like the lingering carts, slowing down as the trail got stronger. That's lovely behaviour for a creature and terrible for a builder, because it crawled exactly where it was supposed to build. Over a simulated month it laid about four cells of pipe. Making it simply faster made things worse: it skimmed past everything. What worked was speed plus trail-following. It cruises along ridges and slows without stalling, and a path now needs two visitors' marks instead of three. Same month: about seventy-five cells.

Then honeycomb. It grew in strands along bee routes, and every attempt to grow it faster just made thicker strands. The fix wasn't how fast comb grows but where it starts. A hive now only takes hold away from other comb, then grows outward, filling the cells it already surrounds first. The routes turned into beads.

Then the two had to meet. Pipe never cuts through comb now. Instead, traffic between two hives counts toward pipe the same way traffic between two marks does, so the routes bees and machines wear between hives harden into pipe. When there's more to build than V-0 can handle, other machines pitch in, and the bigger the backlog, the likelier they are to help. Swarm researchers call that a response threshold (Bonabeau, Theraulaz and Deneubourg, 1996). Hives joined by one pipe network pool half of the pollen they gather, so a busy hive keeps a quiet one alive.

Last, longevity. Bees live about six months now instead of one. Big hives fade slower, and comb that bees keep coming back to hardens into wax that outlasts the rest.

As of the first week of October, about a hundred field days in, the live field has its first pipe and a few hundred cells of honeycomb, built by six bees. None of it was drawn by anyone.