public repos · fifteen months · nights and weekendsdates → repos → writeups
Apr 2025 → now · the main run · one 2024 prologue below
Projects over time
Bars run first commit → last build commit, so overlaps are visible. A few repos kept getting small restyles after their bar ends. Click one to jump to the writeup.
How they were built · each writeup below carries one of these labels
spec-first A PRD and tech spec before any code, with tasks numbered from the spec. The oldest habit here; it started in 2025.
incremental specs No big up-front plan. A dated spec folder per release, then build, tag, and repeat.
ad hoc Conversation-driven, no spec. Some of the quickest builds on this page went this way, and it shows in their scope.
dark factory An unattended harness planned, built, and verified the work. I read the commits in the morning.
now · Jul 2026the latest build
01
Jul 2026 — present · 2 repos
Anima Mesh The company-of-0 experiment
ad hoc no spec preceded the code. The discipline came from releases: 26 tags in 19 days, a changelog per release line, learnings written down as they happened.
The experiment: how much of an organization's back office can a mesh of agents run on
its own? Compliance calendar, bookkeeping, document library, board hygiene, research
watch, inbound triage. The point is to push that boundary until one human job is left
standing: the approval gate.
The engine is public and knows no names. The knowledge lives in a second, private
brain repo: ledger, records, CRM, all plain markdown. One daily heartbeat reads the
repo, runs the agents that are due, verifies the work in code, and commits.
git log is literally "what did the mesh do while I was away."
· engine + brain, born seven minutes apart
One engine, one private brain per instance
The split is the design: all logic in a public engine that never names anyone. All identity in a private instance repo the engine animates.
The cloud tier. Two Durable Objects, one alarm, zero databases. Git is the only durable store.
Hand-authored SVG · from the repo's own docs
The system, in four diagrams
One organization as plain files: knowledge, roster, and the trail the work leaves behind. No code lives here.One alarm, the whole company. Spokes run first; the hub runs last and reads the day's work instead of commanding it.An agent is a markdown file. Hiring is copying a template. Promotion is a one-line frontmatter edit, with git as the record.One-way dependency. The engine never references the brain. One engine can animate many isolated companies.
The beat · 08:00 America/New_York, every day
0 ledger entries0 agent reports0 commits by the machine0 agents on roster
Running since · cloud-native since Jul 11
Nine agents, one approval gate
From the docs: "the initiative inverts: you stop driving the operations and start reviewing them."
Nine agent archetypes on four rhythms: daily, weekly, monthly, quarterly. A compliance
watcher, a bookkeeper, a librarian, a board secretary, a research watch, an inbound
triage. The hub, a Chief of Staff persona, is the single voice the human ever hears.
Three inbound channels: Discord, a dashboard, plain email. And a documented moment of
real judgment: the hub woke the bookkeeper early because a deadline outpaced its
monthly rhythm. That is exactly the judgment the mechanism exists for.
"A reply that isn't backed by a commit is a hallucination with a send button."
26 tagged releases in the first 19 days · ~296 tests · Apache-2.0
Safety lives in code, never in prompts
Everything between wake-up and gate is model judgment. Gates, ladder, ledger, and verifiers are deterministic code.
Five constitutionally gated action classes: money movement, government filings, external
publishing, credential exposure, access expansion. Three commercial agents are fully
written and the engine refuses to run them until their activation gates open.
Capability never outruns permission. And when api.kimi.com's WAF blocked all Cloudflare
Workers egress in week one, cognition rerouted to Anthropic with a two-line config edit.
The agent's identity never changed.
0 databases · git is the store6 model harnesses, swappable3 agents waiting on permission
Feb — Jul 2026 · 2 repos · co-developer: age single digits
Built with Bruce The father-and-son repos
ad hoc he pitches the design, we talk, it ships. No spec survives a co-designer under ten.
Not every project is an agent mesh. Two of the repos are games designed with my son:
he supplies the game design, I supply the AI, and an idea becomes something playable
before bedtime.
Feb 28 — Jun 6, 2026 · 52 commits
Lava Dash
A Geometry-Dash-inspired side-scroller in React and Canvas. One player, then a
two-player co-op mode, because the designer insisted his friends needed to play too.
The same tools that run the mesh also ship a card game before bedtime.
rewind · Jun 2026the third coordinator
03
Oct 2025 — Jun 2026 · 2 repos · 317 commits · 16 tagged releases
The Content Factory A decoupled agent mesh
incremental specs Spec Kit bundles (spec, plan, tasks) in the v1 era; a dated plan in ai-docs per release in the mesh era. Six days from PRD to deployed mesh, because the spec came first.
Talk material for adaptTo() 2026, built for real: a mesh of independently-addressable
A2A agents that generate, migrate, and evaluate CMS content. Different vendors compose
over an open protocol. Quality is measured, not vibed.
The arc runs in three eras. Oct 2025: an MCP server for AEM's da.live, on Azure
Functions. A 94-commit December push tags v1.0 on New Year's Day: evals in a Docker
container, Make.com scenarios calling the Functions to do the authoring, and by
mid-January, Adobe Summit 2026 blog material coming out of the same repo. Then a
quiet spring. In June, v2.0 rebuilds it as four agents: coordinator, content-gen,
migration, eval. Six days from PRD to a deployed mesh on Cloudflare Workers and
Containers.
One pipeline, two model vendors, zero contract changes. Kimi authors. Claude judges. Neither knows the other exists.
The closed loop, from the repo's own docs. The coordinator can enter or exit the loop anywhere.
91 eval score on a real Kimi-authored page~9 min for a real agentic migration~11 min full closed loop, real everything4/4 cloud acceptance, e2e
One eval engine, two deployments
v1.0
A deployable container: a Next.js eval app with four agents, each pairing deterministic checks (axe-core, Cheerio, Playwright) with model judgment. Still running as the frozen backup.
the engine
The scoring engine survived the rewrite. It was copied out of the v1 app, not rebuilt.
v2.0
An addressable A2A agent in the mesh: any vendor's output gets scored the same way, on every run.
The through-line, v1.0 → v2.5 · the biggest single component
The eval outlived both architectures
If agents author content, something has to say whether it's good. That something gets versioned, deployed, and tested like a product of its own.
Four dimensions: structure, accessibility, content fidelity, visual correctness. In
v1 the eval shipped as its own Docker container. In v2 it became an agent in the
mesh, so every migration lands with a score attached. The payoff came in v2.5, when
the numbers said the site redesign was failing: the error turned out to be in the
eval's own methodology, and scoring generated pages on their own merits moved a real
article from 67 to 89. Numbers you can argue with beat vibes you can't.
The demo, before and after
before
"Katie's Travel Journal," a hand-coded, circa-2002 travel blog. Plain HTML, zero framework.
the mesh
An agent mesh migrates it: one vendor authors into da.live, another scores structure, accessibility, fidelity, visuals.
after
The Wilderness Journal: a modern Edge Delivery Services site, 14 blocks, every migration scored.
· v2.2 · the factory runs daily
A daily unattended run
A daily cron submits a run with no topic. The coordinator ideates one, generates, migrates, evaluates, and parks it in preview. A human curates later.
Two lessons came out of production. Streams die crossing the Worker-to-container hop,
so the store, not the stream, is the contract: every agent persists full task state and
consumers recover by polling. And containers get no native database bindings, so the
Worker proxies queries.
20.6K lines of TypeScript27 e2e files · no mocks, ever5 JSON Schema contracts
Five subsystems around one EDS site. Color is the semantic axis this page's diagrams still use: deterministic in blue, agentic in red, mixed in purple.
Drawn while v1 was being built: Agent SDK content generation, a Make.com builder over MCP, the evals container, and the site-building harness. The Excalidraw was later recovered from git history.
agent
AEM
create page · update content · publish, as tool calls
· where this repo started
First, AEM content operations became tool calls
The origin build: an MCP server for AEM's da.live, then the same server proven from Claude, Gemini, Azure OpenAI, and n8n. No client-specific forks.
The repo's first life was interoperability with the CMS I work on professionally:
expose AEM Edge Delivery's content operations, get, update, create, publish, over
MCP once, and every agent that speaks the protocol inherits them. Everything else in
this chapter stands on that server. The mesh grew out of asking what happens when
the agents themselves get addresses too.
Mar — Jun 2026 · 54 commits · a public plugin marketplace
AI Builder Kit Skills anyone can install
ad hoc skill by skill, as the need came up. Each ships docs and guardrails; the heaviest one ships its own evals.
A Claude Code plugin marketplace: nine published skills, written so that people other
than the author can install and use them. Two of them get detail here, because they
carry the most engineering.
Featured skill · Playwright spec → narrated MP4
A test file becomes a demo video
The rule that makes it work: you cannot narrate what doesn't exist.
One command records the spec with exact per-action timestamps, generates ElevenLabs
voiceover from the captions in the test file, and merges freeze-frames and music into
the finished MP4. Before any narration is written, a mandatory stage zero verifies the
app has real data and runs an exploration spec cataloging what Playwright can actually
see. Scenes are built one at a time, each recorded and verified before the next.
The skill's reference output. The four-color semantic palette here is the same one on diagrams across this page.
Featured skill · text → draw.io XML → rendered PNG
Diagrams that check their own work
The skill writes the XML, exports a PNG through the draw.io CLI, then reads the image back and fixes what looks wrong.
Every diagram goes through a render-view-fix loop, usually two to four passes, with
versioned files kept as iteration history. An eight-question interview with recommended
defaults settles the decisions that shape the result: frames, flow direction, color
axis, how exceptions are marked. The style rules live in reference files the skill
must read before generating, because the rules interact.
The skill also ships its own eval harness: two benchmark diagrams, graded by a script
against roughly twenty assertions each, XML mechanics and brand rules alike. The
baseline run without the skill passed 15 of 18 assertions; with the skill, all of
them. The second iteration folded review findings back in as three new assertions per
eval, and the runs are kept in the repo. A skill you can't score is a skill you can't
improve.
The rest of the shelf: the HLX admin skill wraps an admin API so configuration becomes
describe, review the diff, apply. Branded decks, conversation loggers, a structured
git-commit skill, competing-PR triage, and an AI-knowledge harvester round it out.
Each ships with its own docs and guardrails.
May 31 — Jun 5, 2026 · 19 commits · phase 1 complete
Conversion Factory Agents that evolve site blocks
spec-first the PRD and the architecture diagram were committed before the agents existed.
An agent runtime for conversion-rate optimization on Edge Delivery Services. A
coordinator forms a theory from analytics, four spoke agents produce a variant block on
a git worktree, and the change ships as a pull request into a live A/B test. Winners
get promoted into the codebase by a separate PR. A retrospective agent then edits the
agents' own instructions, so the system learns across cycles.
· PRD + architecture diagram, drawn first
Four phases, one diagram
Analyst scores the block. Architect forms the theory. Builder writes the variant. Validator runs Lighthouse, accessibility, and regression.
From the repo's ai-docs. Phase 4 is the part I care about: the retrospective agent proposes instruction changes, and humans gate how the agents reason.
4 spoke agents behind one coordinatorA/B baseline vs variant, per visitor1 false-PASS hole closed: verify wiring, not just pixels
Jan 24 — May 31, 2026 · 374 commits · 16 releases · ran 24/7 for four months
Continuous Executive Agent A 24/7 coding agent
incremental specs to build it: a dated spec and plan per release, ai-docs v1 through v3. dark factory when it ran: nobody watched the overnight builds.
A coding agent that runs around the clock under PM2. It selects work from a priority
queue, spawns isolated workers, validates with deterministic verifiers, retries with a
different strategy on each failure, and escalates to a human only when stuck.
This section goes release by release: a foundation era, an identity-and-multi-vendor
era, and a memory era. Four months. The output repo ended up with thirty projects,
most of them built overnight.
Lifetime, measured from its own ledgers
0 worker contracts0 goals completed autonomously0 projects built0 verifier runs0 human interventions, total
Counted from the append-only ledgers in the repo.
v1.0.0 → v3.0.0 · 16 tags
Release history
Dates from git tags. Summaries from the changelog.
Era 1 · Foundation · Jan 24 — Feb 4
v1.0
Jan 24, 2026 · built in a single day
Foundation
The 8-phase executive loop. Two-repo architecture: the brain never receives worker output. A constitution of 8 immutable limits, including a $20/month spend cap. Contracts, deterministic verifiers, PM2, append-only ledgers.
v1.1
Jan 25, 2026
Intelligence
Goals over 100 turns auto-split into steps. Strategy rotation: every retry must try something different. Agentic diagnosis after three failures. The needs-you.md protocol: I answer in a markdown table, it notices within thirty seconds.
v1.2
Jan 28, 2026
Organization
Goal bundles: each goal a directory with its own prompt, machine-readable steps, contract log, and progress log. Workspace lifecycle drafts → ondeck → in-progress → completed, promoted by priority.
v1.3
Feb 4, 2026
Security
Three physically separated credential tiers with leak detection on every spawn. OAuth-first: the whole system runs on a Claude subscription, no API key.
Era 2 · Identity & Multi-Vendor · Mar 29 — Apr 19
v2.0
Mar 29, 2026
Identity: the agent gets its own accounts
Its own Gmail: it receives goals by email. Its own Discord: it reports completions and blocks. Skills and playbooks with a track record: every skill carries a success rate. A read-only dashboard. I became a guest in my agent's workspace.
v2.1
Apr 1 — 5, 2026
Multi-vendor: the bake-off
A vendor abstraction layer: Claude, OpenAI Codex, and Kimi workers behind one interface. Then the experiment: the same finance-dashboard spec run across all four vendor modes, deployed live, and scored by the harness itself.
v2.2
Apr 12, 2026
Harness framework
Three multi-agent plan-then-build pipelines became first-class citizens, runnable standalone or driven by the executive. 105 passing tests; the whole orchestrator exercised end-to-end in about a second with zero API calls.
v2.3–2.4
Apr 18 — 19, 2026
Build targets & hard lessons
Worktree isolation per project. Progressive skill disclosure with a verifier that fails a contract if a required skill was never consulted. And a favorite war story: a prompt saying "No UI, no database" got post-processed into a plan whose step zero provisioned database tables. The hallucinating prerequisite-inserter was found and fixed.
Era 3 · The Second Brain · May 16 — 31
v3.0
May 31, 2026
Memory
A mem0-backed second brain. Five memory types, five lifecycle hooks, memory packs baked into each worker's context, ~134 memories backfilled. Git stays canonical: if the store and the repo disagree, the repo wins. The riskiest hook stayed off until the memories earned trust.
Same spec · four vendor modes · scored /130 by the harness
OpenAI Codex115
Kimi K2.5 CLI92
Claude88
Kimi Wire87
All four deployed live from the agent's own output repo.
· demoed live at AI Tinkerers Columbus
The multi-vendor bake-off
Same spec, four vendor modes, seven verifiers each. The harness did the scoring.
Demoed on the running system, no slides. One night that week the ledger shows Codex
finishing its build and passing validation at 3:56 AM, with nobody awake. The demo
video shown was produced by the agent end to end: 81 seconds, zero keystrokes.
"Is it perfect? No. Is it close? Also no. But I'm proud of where it's at."
Two weeks later, AI Tinkerers picked the demo as their #1 top pick globally that week and made this graphic. Their design, not mine.
The agent's own Discord. I read it like a colleague's channel.
· the identity era, up close
Its own Gmail, Discord, and Notion
An always-on agent needs its own accounts, not borrowed ones.
It receives goals by email and reports to its own Discord channel. I am a guest in the
agent's workspace, and the working rhythm is reading the overnight commits in the
morning. Human interaction is markdown files, not a chat box waiting for input.
The v2.1 system map, hand-drawn: executive brain, worker execution, feedback loop.
The write path and the read path. One writer, one reader, no agent scribbles wherever it wants.
· v3.0 · the closing release
The agent gets a second brain
Storage is the easy half. Retrieval decides whether a memory was ever real.
Five memory types, from immutable principles to reflective lessons. Workers never touch
the store: they get a curated, two-thousand-token memory pack baked into their context.
The best war story is the silent bug: an empty-string config value made every single
write fail while the logs said "ok." Found, fixed, and the logging redesigned so a total
failure can never look like success again.
Feb — Apr 2026 · 48 commits · live demo + published video
CIAM Demo Customer identity, end to end
spec-first specs and prompt logs kept in the repo's ai-docs.
No agents in this one. Customer identity and access management, built out fully:
identity provider, session boundary, resource API, multi-tenant authorization.
Keycloak
Phase Two · OIDC identity provider
→
Next.js BFF
Auth.js v5 · session boundary
→
Spring Boot API
org-scoped claims · RBAC
Organization-scoped claims flow from Keycloak through a backend-for-frontend into a
Spring Boot API that enforces role-based access per tenant. Deployed on an ARM64 VM,
demoed on video. 48 commits, Feb to Apr 2026.
Jul 2025 · 2 repos · two stacks · restyles into 2026
Postal Portal A shipping product on two stacks
spec-first PRD, then tech spec, then tasks. The transactional flow's commit log counts up from task 1.2 to 9.3.
A postal shipping product built across two weeks of July 2025, spec-first on both
halves. The marketing site went on Adobe Edge Delivery Services as 17+ reusable
authoring blocks. The transactional flow, quote through pickup through confirmation,
went on Next.js: 89 commits in a six-day sprint, each one numbered against the task
list it was working through.
The git history under-counts this one. The marketing repo's whole build week landed
as a single commit touching 2,700+ lines, the PRD and tech spec its README points to
never got committed, and the page content lives in da.live rather than git. Small
restyles kept trailing in afterward; the last one landed in May 2026, ten months
after the sprint.
One spec method, two stacks. Drawn for this page from the two repos' commit history.
May — Jul 2025 · my first fully AI-written production app
Shadow Pivot My first AI-coded app
spec-first where the habit started: PRDs, rules files, blueprints, then code.
A public, production Next.js agent app where AI wrote all of the code. Then CI/CD stood
up around it, and the methodology written down: PRDs, rules files, blueprints.
Three videos documented it as it happened.
Act 1 · Release 1.0 The app ships. 100% AI-written code, one person directing. Post ↗Act 2 · Infrastructure CI/CD and deployment, so the loop can run without ceremony. Post ↗Act 3 · Methodology The repeatable part extracted: PRDs, rules files, blueprints. Post ↗
ad hoc one Sunday, the same app tried in several tools.
The first project: a Sunday spent trying AI coding tools by building the same app in
each. Azure Functions plus Azure OpenAI, turning CSV feedback exports into STAR-format
career stories.
I stopped working on it once the later projects took over. Fifteen months separate
this repo's first commit from Anima Mesh's. Everything above happened in between.
Aug 2023 — Mar 2024 · won at Adobe Summit · repo public Nov 2024
AEM Rockstar 2024 The prologue
ad hoc hand-written Java and jQuery, versioned 1.0 → 1.2 through fall 2023. The AI was the feature, not the author.
The pitch that won AEM Rockstar at Adobe Summit 2024: personalization at scale,
inside AEM's own authoring UI. An author opens an experience fragment, types
"rewrite this for ages 13 to 17," reviews a side-by-side table of every proposed
change, and one click later has a new fragment variation open in the editor.
The design constraint that shaped it: add capability without adding surface.
Generation is a Sling selector on the stock experience-fragment page type, the UI is
one clientlib injected into the stock properties editor, and writes go through
AEM's own copy and POST machinery. No new endpoints, no new editor, no writes
without review.
The whole flow. Drawn for this page from the repo's code.
The frontend is where most of the craft went. A ~270-line Granite/jQuery clientlib
hooks foundation-contentloaded, derives the fragment path
from the properties-editor URL, and builds the review table in the DOM: path,
original value, generated value, one row per property the QueryBuilder walk found.
A continue-chat toggle threads the chat id back, so the model refines its last
answer instead of starting over. Approval clones the variation through
/bin/wcmcommand, writes the accepted values with a Sling
POST, and links straight into the new fragment's editor. A second clientlib does
the same for whole pages.
None of this code was AI-written; in August 2023 the AI was the product, GPT-3.5 by
default with GPT-4 behind a config. The shape it demoed is the one everything above
formalizes later as propose and dispose: the model proposes values, a human reviews
every row, and deterministic machinery does the writing. In 2023 that pattern was a
jQuery table.
0 new endpoints · Sling selectors on stock types2 clientlibs: fragments + pages1.0 → 1.2 releases, Aug — Nov 2023
