The Lossless Group · Field notes

Transformative Highlights from an AI + Agents + Human Collaboration

Notes from a 2 Year Rabbit Hole

Michael Staton

Punched in the face → bucked by the horse

“Everyone has a plan until the AI model punches you in the face.”

— Mike Tyson

  1. Me

    “There’s a new way to develop products now, I want to show you through a demo project.”

  2. Jon Laerdal

    “Can you have AI help our product and design teams understand more about our customers?”

  3. Me

    “Sure!” (How hard could that be?)

Punch punch, then buck.

The New Paradigm was the Old Paradigm within 6 months in 2024.

1 · The old new paradigm

  1. 01Design System
  2. 02Figma clickable paper prototypes
  3. 03Engineering-level documentation
  4. 04Steal from UI kits, use Backend-as-a-Service
  5. =Product Velocity

Coding agents blew this up by Jan 2025.

2 · The new new paradigm

  1. 01Watch agents make weeks of progress in minutes.
  2. 02Get overwhelmed with excitement.
  3. 03Watch agents punch you in the face and buck you off.
  4. 04Tell agents and models you’re quitting and going back to 1.
  5. 05Wake up and try again anyway.

↺ repeat

How fast the ground moved

$60B

Cursor, from launch to a $60B acquisition in about three years.

First release in March 2023. SpaceX announced it would buy Anysphere, Cursor’s parent, on June 16, 2026 and closed the deal on August 14, 2026. The largest acquisition of a venture-backed startup ever.

Feb 2, 2025

“Vibe coding” got its name.

In a tweet Andrej Karpathy later called a “shower of thoughts throwaway.”

51%

More than half of the code committed to GitHub is now written with AI.

Week of April 21, 2026.

My idealism, delusions, and gratitude made for a Wonderland of a rabbit hole.

1 · My intention

Play a meaningful role in catalyzing breakthroughs in Laerdal’s path to thrive for another 100 years.

How? By translating best practices from Silicon Valley, startups, scaleups, and world-class tech giants.

3 · What I hoped

When at Learn Capital, I would come to Stavanger and everyone would make time for me, I would present to a crowded room, have teams assembled to give them feedback.

I was hoping I could quickly repeat that for AI Readiness + Transformation consulting.....

2 · What I heard

  • “Maybe Michael could help us think about what technology platforms are worth adopting?”

    Tore Laerdal

  • “Maybe Michael could help us think about what we can cut or migrate away from in our SaaS license per person spend?”

    Alf Christian

  • “Maybe Michael could help us reduce the drag we have getting in our own way?”

    Anonymous

  • “Maybe Michael can listen to our issues and goals and knock some sense into all the other people who stall our progress unintentionally?”

    Everyone

And…..

And here is

Two years in the rabbit hole, by the numbers

lines of code
1.9M
commits
5.8K
active repos
32
tool indexesIndexes and explainers: over 1.5M lines of code.
1,770
concepts and vocabulary explainers
600+
context filesAnd my own Claude Code plugin that runs them.
1,166

One true monorepo according to the recommendations I would make.

And a harness plugin, markdown flavor, a content development suite, and a pear tree.

Plus….

VC addicts gotta VC.

  1. 1

    Co-Investment Fund: Hypernova

    8 LP commitments, 15 direct investments.

    Investments include Thinking Machines · Harmonic · Ruya · Bruin · Star Catcher

    1

    Impact Foundation: The Water Foundation

    Water Resilience for 1B people by 2040.

  2. 2 + 2

    SPVs on thesis, and syndicate deals

    Aalo Atomics, Chroma. Trusted Router, Glen.

  3. Now

    A future-of-health fund, Humain VC

    Investments into ProfileHealth, CogSci, NextSense, Radicle Health.

Studies

10 Study repos, covering 93 influential codebases.

  • agent-harnesses

    10

    opencode · goose · aider · OpenHands · modelcontextprotocol/python-sdk · continue · autogen · cline · codex · earendil-works/pi

  • content-engines

    4

    astro-big-doc · content-structure · galaxy · mdBook

  • conversational-ui-and-native-shells

    10

    Kaas · openagent · Dive · 5ire · anything-llm · routa · open-vibe · LibreChat · onyx · portaljs

  • data-analytics-specifications-and-standards

    6

    ggsql · parquet-format · datapackage · vega-lite · arrow · observablehq/plot

  • frontend-ui-kits-component-libraries

    10

    shadcn-ui/ui · tailwindcss · ant-design · shadcn-svelte · bits-ui · melt-ui · radix-ui/primitives · open-props · material-ui · primer/primitives

  • markdown-parse-serve-render

    6

    panache · remark · markdoc · markdown-rs · marksman · pandoc

  • memory-layers-for-agents

    18

    mem0 · Parslee-ai/neo · statebench · Martian-Engineering/volt · mempalace · graphiti · delta-Mem · letta · graphify · honcho · OpenViking · hindsight · RetainDB · supermemory · byterover-cli · beads · Understand-Anything · paxm

  • open-specs-and-standards

    14

    12-factor-agents · A2A · agents.md · anthropics/skills · google-labs-code/design.md · frictionlessdata/specs · llms-txt · modelcontextprotocol · OpenSpec · spec-kit · Parslee-ai/mcp-api · get-shit-done · symphony · superpowers

  • sync-and-content-version-control

    9

    jj · seafile · seafile-server · syncthing · restic · kopia · automerge · rclone-ui · git-annex

  • vector-databases

    6

    chroma · qdrant · weaviate · milvus · lancedb · pgvector

What coding agents can be relied on for, and what they can’t

Can be relied on for

Surprising breakthroughs
Scripts and the command line, learning a stack from its docs, outputting docs as HTML/CSS/JS/JSON, finding things on the filesystem faster than any human, and an always-up, always-down companion.
Code & stacks
The canon stack, migrations and upgrades, test coverage, responsive layouts.
Ops & troubleshooting
Command line hell, configuration hell, sysadmin and DevOps.
Docs & prose
Documentation, readable prose.
Data & metadata
Metadata fill-outs, tidying and completing data sets.

Can’t be relied on for

Judgment & honesty
Sycophancy, faking it instead of admitting limits, holding back commentary.
Memory & context
Recent work, context-window limits, multi-session projects.
Discipline
Complex guidelines, overzealous changes, staying on task when stray warnings appear.
Prose
Writing from real evidence, writing for the right audience with a skimmable hierarchy.
Sources & formats
Web search, proprietary formats.

Augment-It

The thing Jon asked for

Now on version 4

Upload a CSV or spreadsheet, fire AI enrichment passes against it, triage the responses, accept the good ones back onto rows, and promote enhanced rows into new canonical record sets.

  1. Version 1

    A Python scripting backend that ran prompts on customer records. What came back: bunk answers, hallucinations, excuses, filler text, and advice on how to find the info I was asking the model to find.

  2. Version 2

    Demo by January 2025

    Bolt.new (and even once in Lovable). A React + Supabase monolith with three LLM providers. Then shelved: I had other pressing matters I needed to line up.

  3. Version 3

    Summer 2025, with Tanuj

    Much better, as a microservices architecture, but it didn’t use the Recommendations.

  4. Version 4

    Today, started March 2026

    40 independently deployable units: a shell, 20 Svelte 5 microfrontends, 12 microservices talking over NATS, and 7 shared packages. Now at semver 4.0.1.x.

It exists to prove our Recommendations (next slide), and I’m still trying to impose them on it.

The honest audit · September 12, 2026

  • Rigorous at the system level: 196 context-v docs, 101 changelog entries.
  • Absent at the component level: 0 of 40 units with their own DESIGN.md, changelog, context-v, or API contract; 14 of 40 with a README.
“Documentation happened where a human sat and thought, and did not happen where an agent generated a working unit and moved on.”

Recommendations for harness engineering

  1. 01Changelog & CI/CD First
  2. 02Context Vigilance First
  3. 03API First
  4. 04UI Kit CI/CD
  5. 05Docs CI/CD

Put those bones in, and harness engineering speeds up over time instead of stalls over time.

Recommendations

…and for culture

  1. 01

    Have the sales team use Plaud. Record everything. Automate ingestion into the corpus.

    Qualify worthy transcripts to go into the User Research Repository.

  2. 02

    Implement a Flexible Hierarchy model inspired by Amazon.

  3. 03

    “Emergent Adoption Systems.”

    Have a Rho/Brex account. A coalition of the willing.

  4. 04

    Designate a “corpora manager.”

  5. 05

    Participate in open standards and specifications bodies.

  6. 06

    Host a competition, and give mad PR for it.

  7. 07

    Department of Crazy Ideas.

    Still relevant. More relevant.

  8. 08

    Host hackathons on looming projects and ideas.

  9. 09

    Designate or recruit a “Stack Engineer.”

    Not a true developer; “beneath them.”

  10. 10

    Designate or recruit a “Data Readiness” owner.

Thank you

Notes from a 2 Year Rabbit Hole

Michael Staton · The Lossless Group

lossless.group

Slide 1 of –