Hyperdrift Turns One: From Hands on the Keyboard to a Crew at the Wheel
A year ago I was the bottleneck by design. Every line, every deploy, every announcement went through my hands, and I was proud of it — Handcrafted. Not factory-made. was the first thing Hyperdrift said about itself. Twelve months later the fleet checks its own health at six in the morning, an LLM judge decides which findings agents may fix alone, and my required input for a whole day fits in three words: approve, deny, redirect.
That isn't a pivot. It's the same promise — tools that respect the person using them — applied to the person building them. Here is the year, shift by shift, with the film, the chart, and a link to everything it names.
The year in numbers
| ~3,000 commits | across 26 repositories since 21 September 2025 |
| 16 live domains | products, POCs and tools under hyperdrift.io and beyond |
| 70 articles | on this blog, every one a claim we had to be able to defend |
| 61 skills | in the shared library every agent inherits before it writes a line |
| 22 public repositories | under github.com/hyperdrift-io, plus 8 packages on npm and PyPI |
| 3 contest entries | shipped in three evenings, all in judging as I write |
| 1 daily loop | that reads the fleet, names the binding constraint per app, and asks for approve, deny or redirect |
Hyperdrift · 21 Sep 2025 → 21 Sep 2026
One year. From hands on the keyboard to a crew at the wheel.
Twelve months, seven shifts, one destination: ai.hyperdrift.io
Before day one
hyperdrift.ioregistered 24 April 2025. Intel’s first commit eight days earlier.- 7 apps · 1 pipeline
Manual coding
Handcrafted. Not factory-made.
First public words on 21 September 2025. By late October: seven apps, three social platforms, one deploy pipeline — every line typed by hand. Revela alone: 1,465 commits.
- 3 Acts
The Manifesto
The apps are not the point. They are the salute.
8 March: HyperCV goes live, the Manifesto goes public, three Acts name the fight — career, knowledge, capital.
- 60 skills
Skills
When the primitives are stable, agents assemble.
10 April: “We Need Better Tools for Orchestrated Development.” 14 April: the workspace repo’s first commit. typerx turns 800 lines of bash into 299 of Python;
hdbecomes the control plane; AGENTS.md files walk up the tree. By June the skill index maintains itself. - 5 days · all tasks SUCCESS
Databricks
A medallion pipeline, end to end, in five days.
23–27 May: Telecom Horizon — bronze, silver, gold on Databricks Free Edition, one
horizon_demojob, all tasks green, one nativeai_query. The same week Orchestra AI is specced — and held back until there is proof. - 4 workflows → 1
Orchestration
Agents propose. The founder confirms. The loop learns.
3 June: Fleet Orchestrator, Concierge, two brains — one grows, one maintains. Its first live finding: the fleet read as broken — 0 signups, 0 visitors. It wasn’t; the labels were. 16 June: The Crew gets its name. 24 June: the Commander’s Read. 3 July: the Bridge is live at fleet.hyperdrift.io.
- 1 weekly brief
Frontier model
We practice at the frontier, then teach what we practiced.
17–18 June: Vanguard ships a weekly frontier brief — new models, new paradigms, ranked against the fleet’s own state. Tiering puts the frontier model where judgment lives and cheaper models everywhere else: 94% off the bill, on paper. 11 July: Next.js purged; own-stack is the standard.
- 3 entries in judging
Judges
Criticals never auto-heal. Everything else, the judge decides.
4 August:
hd foundation heal— an LLM judge reads the autonomy policy and splits every finding into agent-fixable or founder’s call. Telegram retired; the Bridge is the one pane. 28 August: HD steps in front of human judges — Helm to Google’s All Things Agentic, uk.gov Radar to OpenAI’s WebMCP Challenge (5,616 entrants), unanswered to DEV’s Generosity edition. - Op. 01 · 2026
ai.hyperdrift.io
The machine we built to run our fleet became the offer.
Orchestra AI relaunches with real proof — Helm, own-stack, CLI meets UI — and the Traction Partnership: we build for free and validate on traction alone. Partners Tecknuovo · Vodafone3. Sponsored by Databricks.
- human inputs / day ↓
AGI
Not a model release. A day with zero human initiations.
The exit criteria are written: every day, without the founder starting anything, the fleet checks its own contract, diagnoses the binding constraint, acts on a one-tap confirm, and learns. The founder’s only input: approve, deny, redirect. We count the human inputs per day. The number is going down. For the record: “AGI” appears in our code exactly once — a slider in a telecom forecast.
Seven shifts, in order
01 · Manual coding — September 2025
Seven apps, three social platforms, one deploy pipeline, every line typed by hand. Revela alone carries 1,465 commits. The loop was code → reflect → repeat, and it worked right up until the fleet outgrew one pair of hands.
Read: Introducing HyperDrift · HyperDrift Picks Up Speed · the early tools, still open — genesis, keep-alive, peer-dependency-checker, hyper-post, alpha-drift, yield-max, whats-that-again.
02 · The Manifesto — March 2026
8 March: HyperCV goes live, the Manifesto goes public, three Acts name the fight — career, knowledge, capital. The apps are not the point. They are the salute.
Read: A Manifesto · Act 1 — Stand Up for Yourself · Act 2 — Knowledge Is Power · Act 3 — Take Back What Was Taken. Live: Next Role (HyperCV's grown-up name — the rename), Intel, Capital Engine.
03 · Skills — April 2026
The gap was never intelligence; it was primitives. Eight hundred lines of bash became 299 of Python, hd became the control plane, and AGENTS.md files started walking up the directory tree so any agent inherits the rules of the room it's standing in. When the primitives are stable, agents assemble. By June the skill index maintained itself. Today there are sixty-one.
Read: We Need Better Tools for Orchestrated Development · The End of Script Maintenance Hell · Typer Companion · One Command. Any Machine.. Open: typerx on PyPI, typer-companion, nvim-config.
04 · Databricks — May 2026
A five-day detour with a long shadow. Telecom Horizon: a bronze–silver–gold medallion pipeline on Databricks Free Edition, one horizon_demo job, all tasks green, one native ai_query. One pipeline, not a platform migration — but it was the week Orchestra AI got specced, and deliberately held back until there was proof to show.
The same spring produced the agent doctrine the rest of the year ran on: Compounding Agent Swarms, Multi-Tenant Agent Architecture, Agents vs Automation, the MCP server pattern, and Hydra, the multi-agent mesh those articles were written against.
05 · Orchestration — June 2026
On 3 June the machine switched on: a Fleet Orchestrator, a Concierge, two brains — one grows, one maintains. Its first live finding was that the whole fleet was broken. Zero signups. Zero visitors. It wasn't; the event labels were undeclared and one token pointed at the wrong project. We fixed it by hand, on purpose, because the autonomous fix would have fixed the wrong thing. Diagnose before hands is policy now. The Crew got its name on 16 June. The Bridge — the picture above — went live on 3 July.
Read: The Fleet Watches Itself · The Web App That Replaced the Terminal Tab (live at stack-one.hyperdrift.io/logs, source). Also that month: Next Role became an MCP server any agent can call, and hyper-video-mesh put agents on a video timeline (the piece).
06 · Frontier model — June 2026
Vanguard: a weekly brief that scans the frontier — new models, new paradigms — and ranks it against the fleet's own state. We practice at the frontier, then teach what we practiced. Tiering puts the frontier model where judgment lives and cheaper models everywhere else. The Farewell, Next.js series came out of the same weeks; own-stack — five dependencies, types across the wire, pure cascading CSS — is the standard every new Hyperdrift app inherits now.
Read: Farewell, Next.js: Server-Rendered React in Five Dependencies · Part Two: Types All the Way Down · The CDN You Already Own. Open: own-stack. Built on it since: Stillness (A Mirror That Never Shows Your Face, source), Together (Build the First Room Before You Build the App, source), WakeJam.
07 · Judges — August 2026
Two kinds. First the machine got one: hd foundation heal runs the doctors, then an LLM judge reads the autonomy policy and sorts every finding into agent-fixable or founder's call. Criticals never auto-heal. Telegram retired the same day; the Bridge is the one pane.
Then we walked in front of human judges. Three entries in three evenings, all still in judging — no trophies to report, only that the fleet now ships a contest entry in a night:
| Entry | Contest | See it |
|---|---|---|
| Helm — an agent crew at the wheel of a live four-app fleet, scoped authority per agent, a self-healing drill against a prompt-injected sandbox | Google All Things Agentic (Devpost) | film, 3:41 · source |
| uk.gov Radar — a page that offers its own tools to whatever agent the founder brings, over WebMCP | OpenAI WebMCP Challenge (Devpost, 5,616 entrants) | film, 2:08 · live · source · The Agent Is the Session |
| unanswered — help-wanted issues nobody replied to, matched to what you know, first reply drafted | DEV Weekend Challenge: Generosity Edition (the challenge) | live · source · the piece |
Read: Your Error Page Is a Prompt — what the Helm drill taught about telemetry as an attack surface. A fourth entry, Standup, is being built in the open for Amazon's Agents for Humans as I write.
08 · ai.hyperdrift.io — 31 August 2026
The machine we built to run our own fleet became the offer. Orchestra AI relaunched with real proof — Helm, own-stack, CLI meets UI — and the Traction Partnership: for a few products a quarter we build the automation for free, and the partnership stands or falls on traction alone. Partners Tecknuovo and Vodafone3; sponsored by Databricks. The detour in May turned out to be a road.
Its first delivery is public too: deputy-workforce-mcp, a read-only workforce MCP built on @hyperdrift-io/mcp-service-kit — the same kit behind MCP Maker.
Open by default
The products that hold customer data stay private. Everything we learn building them goes out the door as code other people can run. That was the stance on day one — homecooked software, community-owned tooling — and it's the stance of Orchestra's contract today: you own the code, in your repo, under your licence.
What that looks like after a year, in links:
- Stacks and patterns — own-stack (server-rendered React in five dependencies), @yannvr/auth (passkeys as the new norm), @yannvr/billing, @yannvr/analytics.
- Agent surfaces — nextrole-mcp (remote at
nextrole.site/api/mcp), deputy-workforce-mcp, @hyperdrift-io/mcp-service-kit, uk-ai-radar (WebMCP), helm. - Operator tooling — typerx, typer-companion, hyper-post, nvim-config, peer-dependency-checker, keep-alive.
- Products built in the open — stillness, together, unanswered, standup, stack-one, hyper-video-mesh, genesis.
Twenty-two public repositories, eight packages, three MCP servers you can call today. The full list is at github.com/hyperdrift-io; the one that matters most this month is the one you'd find useful — tell me which.
And then, AGI
I won't tell you AGI is coming. I'll tell you how we'd know it arrived here: a day where nobody starts anything. The exit criteria are written down — the fleet checks its own data contract, diagnoses the binding constraint per app, acts on a one-tap confirm, learns from the outcome — and the founder's only input is approve, deny, redirect. We count the human inputs per day. The number is going down.
For honesty's sake: the string AGI appears in our codebase exactly once, as a slider in a telecom forecast. We're keeping it.
What a year taught
The bottleneck moves; it doesn't disappear. Every shift moved mine — from typing, to writing primitives, to confirming, to deciding what a judge may decide. That's the job now, and it's a better one.
Diagnose before hands. The first thing the autonomous fleet told us was wrong, and the wrong fix was one keypress away.
Ship the machine you'd sell. Every client conversation at Orchestra starts with a dashboard we look at ourselves every morning.
The full chart lives at hyperdrift.io/one. If you run a fleet of your own and want a crew at the wheel, the door is ai.hyperdrift.io. If one of the repositories above saves you a week, that's the salute working. And if you'd rather watch year two unfold, the newsletter is below.
The ledger — every claim above, dated and sourced
| Date | Milestone | Source |
|---|---|---|
| 2025-04-24 | hyperdrift.io registered; Intel's first commit eight days earlier | whois · apps/intel git log |
| 2025-09-21 | First published article | Intel: AI-Powered News Intelligence |
| 2025-10-19 | "Introducing HyperDrift" — Handcrafted. Not factory-made. | article |
| 2025-10-26 | Seven apps, three platforms, one pipeline | article |
| 2026-03-08 | HyperCV live · the Manifesto · Acts 1–3 | launch · manifesto |
| 2026-04-10 | "We Need Better Tools for Orchestrated Development" | article |
| 2026-04-14 | Workspace repo first commit ca4c418 | root git log |
| 2026-04-17 | typerx: 800 lines of bash → 299 of Python | article · PyPI |
| 2026-04-28 | hd CLI + TUI | 37cb71f |
| 2026-05-17 | Orchestra AI specced | docs/superpowers/specs/2026-05-17-orchestra-ai-design.md |
| 2026-05-23 → 27 | Telecom Horizon medallion pipeline on Databricks Free Edition — horizon_demo, all tasks SUCCESS | apps/poc/horizon/docs/DEMO-databricks.md |
| 2026-06-02 | Self-maintaining skill index (#5) | 8e53c5d |
| 2026-06-03 | Fleet Orchestrator + Concierge (#11) · two brains (#16, #17) · first live finding | meta/AI-NATIVE-ORG.md |
| 2026-06-09 | Next Role becomes an MCP server | nextrole-mcp |
| 2026-06-16 | The Crew named | 880aa3e |
| 2026-06-17 → 18 | Vanguard — weekly frontier brief | a6cf841 |
| 2026-06-24 | Commander's Read · commander-daily skill | root git log (13 commits) |
| 2026-07-03 | The Bridge live at fleet.hyperdrift.io | docs/the-crew/ROADMAP.md · The Fleet Watches Itself |
| 2026-07-11 | Next.js → purge; own-stack standard | meta/TOOLING.md · own-stack |
| 2026-08-04 → 05 | hd foundation heal with LLM judge; Telegram retired | docs/superpowers/specs/2026-08-04-hd-self-heal-judge.md · 847ec07 |
| 2026-08-28 | Contest skill + ledger | 6f4ad25 |
| 2026-08-31 | Helm submitted · ai.hyperdrift.io relaunch | Devpost · ai.hyperdrift.io |
| 2026-09-02 | uk.gov Radar submitted (5,616 participants) | Devpost |
| 2026-09-06 | unanswered submitted | dev.to |
Counts as of 12 September 2026: commits are git rev-list --count --since=2025-09-21 summed over the 26 repositories in the workspace; domains from the fleet registry; articles are the non-daily files in content/blog; skills are the entries in the shared skill library; public repositories and packages as listed on GitHub, npm and PyPI. The 94% cost figure in station 06 is the Hydra cost model from Compounding Agent Swarms. Contest outcomes: none yet.
Get HyperDrift signal — courtesy of intel.hyperdrift.io