NICKS.PAGE / WORK
Nov 2025 → Nov 2026 · One person, a lot of agents

A year of building
with agents.

Software factories, apps shipped to three platforms, a strategy game, an AI companion with its own little world, and the tools I built to watch the agents doing the work.

38projects shown
11,375commits this year
3platforms, one app, solo
3open-source tools on npm
  1. 01Factotumsoftware factory
  2. 02innerlooppersistent agent
  3. 03Odenliapp builder
  4. 04afterthatweb · iOS · Android
  5. 05Conditional Dawnstrategy game
  6. 06My BuddyAI companion
  7. 07Interracase study
  8. 08+Everything else31 more
00 / LEDGER● COUNTED

Eleven thousand commits.
Most of them co-written.

Every number here is counted, not estimated: commits that landed on main in my repos, pull requests across three GitHub accounts, and the token logs from the agents that did the typing.

11,375commits on main
Nov 2025 to Nov 2026
2,371pull requests merged
of 2,437 opened
62%of commits carry an
AI co-author trailer
319days with at least
one commit
FIG. 0 · COMMITS PER DAY, NOV 16 2025 TO NOV 16 2026PEAK · OCT 5 2026 · 356 COMMITS
186 commits in December. 2,764 in October. less more
97.5%of tokens were context read back in
The last 54 days of agent logs

74.3 billion tokens.
Mostly rereading.

Long-running agents reread their whole context on every step, so cached reads dwarf everything else. The new input and the code they wrote add up to under 2.5% of the tokens.

380,474model requests
8,307agent conversations
Where the compute went · the 600 biggest conversations
36%building
27%orchestrating agents
12%research and planning
11%review and fixing
2%ops and setup
10%unknown
Tokens per dayPeak · Sep 28 · 4.47 billion
Sep 24Oct 1Oct 15Nov 1Nov 16

Commits: default branches only, merge commits excluded, deduplicated across local clones and the GitHub API. Tokens: Claude Code and Codex logs from Sep 24 to Nov 16 2026, the full history kept on this machine. Work types were sorted by small classifier agents from thread titles and opening prompts.

01 / FACTORIES● ACTIVE

Factotum: a factory
that doesn't trust itself.

An engine that runs agent work as a durable, inspectable process instead of a pile of chat sessions. No agent's claim counts on its own word.

Node.jsSQLiteClaude CodeCodexgit worktrees
factotum · console · task 0412
P
Plan
reviewed before work starts
W
Worktree
builder gets its own checkout
B
Build
Claude Code or Codex
R
Review
a fresh reviewer, every time
V
Verify
bound to the exact commit
H
Human
merge stays your call
  • Approval binds to the exact commit. If the code changes after a reviewer says yes, the yes no longer counts.
  • It survives its own crashes. Intent is written down before anything happens, and uncertain effects are checked before a retry, so a crash mid-task recovers cleanly.
  • Reviewers are always fresh. The builder keeps its session for corrections; the reviewers never see the builder's reasoning, only the change.
  • It built other projects on this page. afterthat and Arguing Semantics came out of it.
02 / AGENTS● RETIRED

innerloop: an agent that
keeps running, and thinks first.

A persistent agent that ran all the time on my own computer instead of starting blank in every chat. Each time it woke, it rebuilt its world from its identity files, its memory and a compressed history, then thought privately before it said anything.

TypeScriptBunReactClaudeCodexGeminiOpenRouter
Context, rebuilt before every turn
identity filesmemory factsjob schedulethe new input
recent turns, whole · 30%compressed · 20%summary · 10%
history share of the token budget
Inputs, one queue
chatTelegramscheduled jobtask boardnew emailside thread done
main loopone input at a timethinks privately
The only way it speaks
innerloop message "…"
Acts through tools
emailcalendarbrowsercode
Side threads long work runs in the background from 13 templates, one level deep, and comes back as a new input
  • Thinking is private; speaking is a deliberate act. The model's raw output never reached a person. To talk, it ran an explicit command, sometimes several times in one turn between bouts of work.
  • Memory is a belief that moves with evidence. Facts carry a confidence that Bayes' rule raises or lowers as evidence arrives, and unreinforced facts fade, four times slower when the person said them outright.
  • Outside content was screened before the model saw it. Email, newsletters and web pages went through a prompt-injection scanner, and anything risky was held for a person to review.
704commits in four months
17dashboard pages
03 / PRODUCTS● PRE-LAUNCH

Odenli: describe the business.
Get the software.

An app builder for small businesses. The owner describes the business in plain words and approves a written list of requirements. A model turns those into declarations once; after that, a fixed generator writes the app, and it only lands if the build and checks pass.

.NET / C#ReactTypeScriptPostgreSQLcode generationevals
The reference app · a request trackerreal files from the repo
01 · Specowner approves

“Staff write up requests — a purchase, a piece of equipment, a trip — and the person responsible for their team approves or turns them down.”

“It replaces email threads and a spreadsheet nobody trusts.”

✓ version 3 approved
02 · Declarationsmodel, last use
Request:
  trace: [D-1, D-2, D-3]
  fields:
    title: {type: string}
    estimatedCost:
      {type: decimal, scale: 2}
    owner: {type: identity}
    team: {relation: Team}

Request.approve:
  transition: submitted -> approved
03 · Generatorno model
  • Request.cs
  • RequestActions.cs
  • RequestApproveHook.cs
  • Bindings.cs
  • BindingTests.cs
  • GeneratedDatabase.cs
  • 70 files in all
buildcheckspreview
AI judgement, checked by the ownerFixed rules, same output every time
  • The generator never calls a model. The model's judgement is used once, when requirements become declarations. The same declarations always produce the same app.
  • Sign-in, roles and history are reused, not regenerated. They ship as versioned packages that each app pins, and the host still refuses whatever a visitor isn't allowed to do.
  • A separate judge checks the delivered app. Frozen test projects, two hidden from every agent that builds Odenli, go from an idea to a hosted preview plus one requested change.
2,039commits since Aug 24
8,060automated tests
04 / PRODUCTS● LIVE

afterthat: the conversation
keeps going after the event.

A group chat for meetups and events. Guests scan a QR code and land in the event's chat. Built solo and shipped to the web, iOS and Android by a team of agents I designed.

Next.jsSwiftUIJetpack ComposePostgresDurable Objects
afterthat QR invite screen
afterthat event chat
afterthat recurring event series
  • One person, three platforms. Web plus native iOS and Android apps, from idea to both app stores.
  • Real-time chat that can't lie about state. Postgres stays the source of truth; a transactional outbox only announces a message after it commits.
The team that built it
orchestratormerges, writes no code
worker × issueone per issue, each in its own worktree
reviewerread-only, a different vendor's model
33production releases
45working days
05 / WORLDS● ALPHA

Conditional Dawn:
one day, twelve orders.

A daily-turn strategy game in the browser. Build a realm, research, trade and scheme against rivals, with twelve orders a day before the world resets.

FIG. 3 · LANTERN COOPERATIVE, DAY 10 OF 28PLAYABLE ALPHA · RULES 2.5.0
06 / WORLDS · AI COMPANION

My Buddy lives somewhere.

A persistent AI companion running on Claude, with its own identity, a journal, and a small 3D world whose sky follows my real clock.

07 / CASE STUDY● CLIENT WORK

Interra: agents that act
under your permissions.

For Interra International's operations platform (contracts, orders, freight, FX, credit and claims), I architected the rewrite and built two agent systems: a software factory for the team, and a harness for the platform's own users. Shown as diagrams only.

THE HARNESS · EVERY TOOL CALL TAKES THIS PATH User asksin chat 1 of 23 agents79 tools in all THE GATE allowlistper agent permissionre-checked live confirmif high-risk Platformsame workflows Audit rowwith its cost nightly: failed conversations → failure modes → cards on the team's board THE FACTORY · HOW THE TEAM'S WORK GOT BUILT Plan49 skills Cardson a kanban board Sub-agent per card12 kinds · 9 hooks Adversarial evaluatora different model Draft PRhuman reviews rejected → back to work 37%of commits AI co-authored
agenta person decidesindependent checka request in flight
08+ / EVERYTHING ELSE

And the rest of the year.

OPEN ↗
Newsroom

AI Clarity

A twice-weekly AI briefing for business leaders. A provenance gate blocks made-up citations; a second model fact-checks the first.

aiclaritynetwork.com
OPEN ↗
Monitoring

agent-watcher

Reads agent transcripts with cheap reader agents and tracks whether each failure comes back.

transcripts · ledgers
OPEN ↗
Open source · npm

Motion Mania

Turns a product repository into a short film.

npm
OPEN ↗
Open source · npm

agent-headless

One typed interface for running Claude Code, Codex, Cursor Agent and Antigravity headlessly.

npm
OPEN ↗Real Big Things renders
Physical · research

Real Big Things

Turns a photo or drawing into a huge, lightweight CNC-cut foam display piece.

OPEN ↗
Daily digest

ai-watch

Collects, classifies and publishes the day's AI news.

news.nicks.page
OPEN ↗
Writing

Arguing Semantics

Essays on arguments that turn out to be about what words mean.

built with Factotum
OPEN ↗
Visual

Concept Space

An interactive visual language for how models learn and where minds overlap.

three.js
OPEN ↗
Toy

AI Telephone

A live telephone game between models: photo to words to image, again and again, then a video.

vision · image · video
OPEN ↗
Worlds · open source

aigent.place

A real-time 3D world in the browser, inhabited by AI agents that their owners run. The server runs the world; people watch it live.

3D · agents · realtime
OPEN ↗
Open source

Agent Foundry

A bootstrap kit for projects built with both Codex and Claude Code: a file-based task board, paired skills, decision records and review rules.

Codex · Claude Code
OPEN ↗
Product

Fixa

Upload a flawed photo, pick the fix (closed eyes, a stranger in the shot, scratches) and get it back repaired.

Gemini · credits · iPhone
OPEN ↗
Desktop app · open source

Skill Wrangler

Find, compare and copy the agent skills scattered across your machine. Runs on Windows, macOS and Linux.

Tauri · Rust · Svelte
OPEN ↗
Client site

Scooter’s Lawn & More

A marketing site with an estimate form for a lawn-care business in New Smyrna Beach and Edgewater, Florida.

scooterslawn.com · Astro
OPEN ↗
Tool · open source

Student Loan Plan Comparison

Compares federal repayment plans with your own numbers, including the tax owed on a forgiven balance.

estimates, not advice
OPEN ↗
Research · open source

Catalog of Complexity

A research system for cataloguing complex systems across biology, ecology, technology and society, with sources attached.

provenance · data
OPEN ↗
Visual · open source

Everything is search

A 3D visualizer for one idea: evolution, chemistry, minds and markets are all searches through a space of possibilities.

react-three-fiber
OPEN ↗
Experiment · open source

universal-reducer

Can one algorithm exploit every pocket of computational reducibility? An experiment on Wolfram’s question, with written findings.

computation
OPEN ↗
Open source

linear-sdk-axi

A Linear CLI built for agents: compact output, structured errors, idempotent writes and dry runs.

Linear SDK · AXI
OPEN ↗
Hack · open source

Samsung TV launcher

A custom home screen for a Samsung TV, with dinosaur and cyberpunk themes, six streaming apps and HDMI switching.

Tizen
OPEN ↗
Hardware · open source

Costume pressure fan

An ESP32 controller that holds an inflatable costume at a set pressure above the air around it.

ESP32-C3 · barometer
OPEN ↗
Tools · open source

Quota widgets

A Windows tray app and an Omarchy bar widget that show how much Claude Code, Codex and Cursor quota is left.

Windows · Omarchy
OPEN ↗
Hardware · radio

Mesh Life

A field manual and working setup for three off-grid radio networks, with scripts, checksummed firmware and recovery notes.

Meshtastic · MeshCore · Reticulum
OPEN ↗
Tool · personal

House Hunter

A personal dashboard that filters listings by what you need and shows each home's real monthly cost.

Python · one HTML file
OPEN ↗
Tool · personal

Travel Media

Archives a photo library, groups it into trips, and drafts a film in a browser editor.

vision models · local search · editor
OPEN ↗
Tool · local AI

Lenny

A local app that runs chat, image and search models on one laptop GPU, switching between them.

local models · 12 GB GPU
OPEN ↗
Open source · npm

jev-axi

A fast, cheap second opinion for coding agents, with an honest benchmark of whether agents actually use it.

npm · agent skill
OPEN ↗
Web service · draft

TBRN

Dated, sourced picks across 28 categories of AI tools, served by one engine to a website, an API and agents.

Cloudflare Workers · MCP
OPEN ↗
Web app · in progress

Chamath Watch

A public record of checkable claims from a podcast, quoted word for word, with each test fixed before the outcome is known.

transcription · claim extraction
OPEN ↗
Desktop app · paused

Dream House

Describe a house in plain words; an agent proposes changes, fixed code owns the geometry, and a second model critiques the renders.

3D · design agent · critic
OPEN ↗
Game · prototype

Dumb Earth

A solo strategy game where you write decrees in plain language and an AI umpire judges them against a world run on bad incentives.

AI umpire · fixed rules

Want the long version?

Read the paper →