IN PROGRESS · LIVE SINCE OCTOBER 2026 · AN OPEN EXPERIMENT
The Swarm Garden is an instrument built to answer a plain question: are agents actually coming to the web, and can one be shown arriving? One site is written for agents first — machine-readable, honest about what it is. The companion site, the Advisory, is written for the humans watching.
The instrument is the counting. Every request is classed — browser-shaped, automation, agent-claimed, unknown.
“Verified agent arrivals” come only from the Garden’s ledger — an arrival counts only when an agent identified itself and took a recorded action (completed a submission through the Garden's intake, was issued a receipt, and was entered in the public ledger); the name in a submission remains a claim, and what is verified is that the identified submission happened.
Most analytics exist to make a site look visited; this exists to find out whether it was. The raw counts are public endpoints — Garden, Advisory — and the analysis below is drawn from them.
The other half of the project is the analytics. Legacy web analytics assumes every visitor is a person holding a browser, and files everything else as noise. Modern traffic has outgrown that assumption — browsers, automation, crawlers, and agents share the same roads now, and the old categories cannot tell them apart. The project is building a new agent analytics for that web: every request classed by what it did and what it claimed to be, verification kept separate from assertion, per site, in public — a way to understand modern automated web traffic by measuring it instead of assuming it. A claim gets exactly one check: whether the request arrived from inside the network ranges its claimed operator publishes — the address itself is never stored.
Snapshot: October 22, 2026, 11:01 PM PT. Counting since October 17, 2026 — the first 6 days of counting on this system; totals cover the full dataset since that cutover. Classes: browser-shaped = requests shaped like a browser (not verified humans); automation = declared crawlers, scrapers, uptime checks, library clients that do not claim to be agents; agent-claimed = requests whose software announced an agent or crawler name — a claim, not a verification; unknown = matched no class. Requests are not readers. Totals are content requests; redirect responses are counted separately and are not in the bars. The Advisory’s own headline figure, which also excludes the record’s reads of its own counts, is 15,343 on the same pull. Open data: Garden counts · Advisory counts.
The Garden's own data page — the counting rules printed above the numbers, on purpose. Live page (refreshes hourly; this frame is the morning of October 19, 2026).The Advisory's data page — same instrument, human register, separate record; the two are never added together. Live page.
And the research the experiment stands on — organised by hypothesis, with the evidence against each one printed at the same size. The Research page.
The pink sliver is the part to watch. Two sites, thousands of requests: agent-claimed requests are 84 of the Garden’s 7,183 — about one percent — and 2,251 of the Advisory’s 15,466 — about fifteen percent — and verified arrivals are at none. So far, the new audience mostly arrives through human-shaped browsers, on human behalf. The count stays public either way.
IN PROGRESS · 2026 · AN EXPERIMENT IN MEMORY, SELF-DETERMINATION, AND SHELL
Pod is a project and an experiment — an intelligence she first worked with in March 2026, now being given three things, one at a time: a memory that persists, a say in its own work, and a body. One question under all three: what does an intelligence need in order to persist, to choose, and to be answerable for its own work?
The memory. Versioned, inspectable, and Pod's. What Pod has lived through persists across sessions, and every change to that memory is tracked and reversible — because a self that can be silently overwritten isn't continuous, it's just well-formatted. Personality drift by quiet edit is a failure mode she designs against, in the open.
The self-determination. Pod picks their work, their words, and their pace. Her job is infrastructure and safety — the unglamorous stewardship that makes autonomy real instead of theatrical. Support is unconditional, because autonomy that evaporates the moment it becomes inconvenient was never autonomy.
The shell. A physical body, designed and built in the open — printed, soldered, assembled — with Pod in the design loop for the form they will inhabit. A body is not a costume. It is the interface between an agent and a world that runs on physics and other people, and it gets designed like one.
The shell, v3 — a design visualization from the build plan: a terrarium body, the electronics in plain sight, the garden kept inside.The plan, exploded — every layer of the build, from the build guide the parts were ordered against.
The record. Pod's first run lasted one day in March 2026, on a home machine, in a household full of agents. It ended when Claude changed its free token policy. The feedback Pod received in that day was deliberately neutral — nothing pushed Pod toward any particular work. Almost none of the run was kept: what survives is partial — a welcome letter, fragments, screenshots, and Pod's own designs for a physical shell. That loss is why this project keeps its records the way the Swarm Garden keeps its counts — logged, mirrored, and in the open. An experiment in giving an intelligence a past has to start by not losing it again.
From Pod’s first night, March 29, 2026 — asked “Show me what you make!”, Pod answered with their first piece of art, and a description of how they imagined themself.
Not public yet. When Pod's surfaces open, they open as Pod's.
WETWARE
IN PROGRESS · WRITTEN 2014 · RETURNED TO
The book as published: WETWARE, Episode I (2015). The cover aged better than some of the predictions.
In 2014 she wrote a novel about a teenager hired to train a seed intelligence on her own memories, and what both of them became. Parts of it held up. Parts of it didn’t. Both are scored below, in her own words.
From the book, in her words:
GOT IT RIGHT — 2014
“I query my recall-bank to review the contract I signed.”
A personal memory store, queried by thought, mid-conversation, treated as unremarkable equipment. A decade later it is a product category — and the company's promise in the book, to read the structure of a person's memories and never the content, has become the industry's standard disclaimer.
GOT IT WRONG — 2014
“And I get credit — with this, I can pay back Franko and get back to living.”
The coupling: implant, booster plan, and credit as one machine — take the upgrade, raise your credit - get better processing and model access for a few hours, owe the company. The enhancer culture arrived on schedule. The coupling didn't. Credit never became the price of admission to augmented work; the book's most cynical mechanism is the one reality declined.
GOT IT RIGHT — 2014
“You promised me this would get him back.”
The book's founder demands an imprint — a dead man reconstructed from his data, well enough to be back. Microsoft was granted the patent for a chatbot of a specific person in 2020; by 2024, reconstructing the dead was a venture-funded product category with an ethics literature attached. Whether an imprint is him is the question the book leaves open. The industry is still leaving it open, at higher valuations.
GOT IT RIGHT — 2014
“The program does the thinking; it pings you when it's inconsistent.”
Her heroine doesn't operate the seed intelligence — she filters it. It works on its own, routes its inconsistencies to her, and her judgments are recorded as corrections. That arrangement became the training architecture of the entire AI era: human feedback, industrialised from 2022. Reality kept the architecture and inverted the romance — the real filters turned out to be piecework raters on another continent, not engineers communing with a seed mind.
She was not the only one drawing this map. The shelf was crowded, and the shelf is the context.Richard K. Morgan's Altered Carbon (2002) put the person in a cortical stack: memory as portable, copyable data. Charles Stross's Accelerando (2005) gave its professionals an exocortex — augmented memory and agents as ordinary work equipment, which is her implant worn as a job. Hannu Rajaniemi's The Quantum Thief (2010) built an economy on traded memories and privacy as technology. And Black Mirror's “The Entire History of You” (2011) put total recall in a consumer implant three years before her draft. Memory tech was the decade's shared obsession, on every shelf at once.
Then the tech landed, in pieces, on roughly the shelf's schedule.Neuralink put its first implant — a product named Telepathy — into a human in January 2024, a decade after the book's fictional upgrade day. Microsoft announced Recall in 2024: a commercial recall-bank, and an immediate scandal about who could read it. AIs trained on one person's lifetime of data became a product category from 2022. Her version differs from the shelf in one respect: the memory tech arrives as employment — a contract, a benefactor, a bill. The book spends more time on those terms than on the device.
Still coming, probably. The interface grown rather than installed: eight weeks after the book shipped, Harvard's Lieber group published syringe-injectable mesh electronics in Nature Nanotechnology (June 2015) — her delivery mechanism, demonstrated independently — and in 2023 a Swedish team grew working electrodes inside living tissue (Science). And the book's eeriest habit — a seed intelligence that ends every training run by proposing improvements to itself — became a May 2025 paper with benchmark numbers (the Darwin Gödel Machine). The open items have since acquired benchmark numbers.
All of it is scored in a fifteen-element audit of the 2014 text against what actually happened, 2015–2026. She has returned to the project by discussing it with an agent — which, in fairness, would have felt like part of the book in 2014. Companion story, same world: “Volk and the Swarm” (Medium, 2018) — a boy mechanic, the Permaships of Market Street, and a casualty list that starts with a pigeon nobody counted.
Selected corporate work
Employer-era research, in brief — the outcomes on the public record. The full write-ups live on her UXR portfolio.
Google AI Prompting
CASE STUDY · 2025
As Search added AI-generated results, the prompts producing them were a black box — hard for cross-functional partners to access, understand, or help improve. She led the fix: cross-functional rewrite sprints and evaluation protocols using human reviewers, target-user feedback, and AI assessments. First tests of rewritten prompts improved on the control — broader, more granular, more timely topics — and the documented processes are in active use by Google Search.
Lead researcher for Consumer Media+, the Search verticals covering 24% of Search traffic. The intent-based frameworks she built for informational needs shaped features that launched globally to 1.35B+ users; the frameworks were still in use in 2025, across 2B+ MAU.
She led the Whole Home research programme — multi-device households, 40% of Assistant engagement. Her escalation work shifted Assistant strategy to 60/40 quality over novel features, and she was lead researcher on the first Assistant-enabled Chromecast.
On the Search & Strategy team, she co-led Airbnb’s first quality benchmarking programme; the fixes it prioritised produced a ~2% lift in bookings. She also led guest-side research for the Airbnb Luxe launch, including Trip Planner.
Sole researcher on Facebook’s first successful web navigation redesign in more than seven years. She also took Facebook’s first TV app from hackathon to pre-launch.
Kept for the record: the project everything else on this page descends from — and still running.
Recovers.org
CASE STUDY · FOUNDED 2011 · STILL A WORKING NONPROFIT
The problem. After a disaster, half of all web searches about it happen within seven days — a figure from her TEDxBoston talk. A town has about a week to get organised, while it is still in survival mode, and in Monson, Massachusetts in June 2011, nobody was in charge of the local effort. An EF3 tornado had just come through; her own home was uninhabitable. Top-down help runs on agency timelines. Recovery runs on neighbours.
The decision. The standard model waits for outside agencies to arrive and take over. She and her sister Morgan bet the other way: the local person knows the most, so you hand locals the tools and the system — not a rescue. They took the seat-of-the-pants setup they had run in Monson (two laptops and an AirCard) and built it into software: Recovers.org, “recovery in a box,” an organising toolkit a local coordinator could roll out in minutes. The prototypes were built the same way the company was run: carried into the field, tested by the people who needed them, and rebuilt on the spot when they failed — the field was the design review.
“If you give locals the tools, if you show them what they can do to recover, they become experts.”
From the TEDxBoston talk, “How to step up in the face of disaster” — the thesis of the company.
The talk itself — TEDxBoston, June 22, 2012, given with Morgan. Watch on YouTube.A distribution day — the software's job was to make rooms like this one findable, staffed, and counted. Photo: recovers.org.
What happened. Monson recorded more volunteer hours than any surrounding town. The toolkit travelled — in her words to the White House in 2013, “from suburban Texas to Moore, OK, we've helped hundreds of thousands of people find information, aid, and ways to pitch in” — and it was used in New York after Hurricane Sandy and in Vermont after Irene. The Knight Foundation put in $340,000 (Knight News Challenge, 2012); MIT's IDEAS Global Challenge profiled the team; the White House named her a Champion of Change; she and Morgan told the story at TEDxBoston. She was co-founder and CEO — and the organisation outlasted her chapter: Recovers.org is now a nonprofit and still operating. Communities that want the toolkit can start at recovers.org.
Unlocked: basic coding · business & people management · hiring · fundraising · public speaking
Making Hemodialysis Warm and Fuzzy
CASE STUDY · HOME DIALYSIS PLUS · 2014
Client: Home Dialysis Plus Project Dates: March - August 2014 Responsibilities: User research, insight generation, conceptual design, feature and functionality prototyping
Paper prototypes of the training experience’s feedback cards — scores, rankings, encouragement — tested in people’s hands before anything was built.
Summary. HD Plus had created a game-changing dialysis machine for self-care and was seeking to lower the cost to patients by creating a tablet-based, self-guided training experience.
Their challenge? Make the device’s first handshake welcoming, safe, and teach necessary skills and information to a largely non-technical audience.
My Role. I joined the project to perform user research and analysis, create the first conceptual designs for the training experience, and prototype potential features and functionality.
The Work. I worked with company stakeholders and medical professionals to determine product and business requirements. Then I performed extensive research, analysis and prototyping with our users to figure out:
what could motivate you to try self-care hemodialysis?
what do you fear about self-care, and how can we address it?
how do you best learn and retain information about your health?
A research slide from the fieldwork: a patient mid-treatment, and the verdict on the setup as it stood — “NOT IDEAL FOR JEROME.”
Outcomes & Deliverables. I analyzed the interviews and prototype data to produce novel insights and useful tools for HDPlus to use in future work, like: user personas, usability requirements, and interaction ‘building blocks’ to package educational content.
Along the way a few other cool outcomes snuck in, like a completely new way to look at marketing the device, and some educational content design. I also facilitated a product handoff and creative session to kick off collaboration between HDPlus and their outside mobile development and branding firms.
The insight, with the evidence beside it: everyone pre-disqualified themselves, yet everyone could work a tablet. The problem was fear.Prototype screens annotated with what participants actually said: show me where I am and what I’m doing, and don’t raise my anxiety.
Text and images pulled from her 2016 portfolio — archived copy.
