+-----------------------------------------------------------------+
| SYSTEM MAP may 2026 -> oct 2026 |
| six months of building things that build things |
+-----------------------------------------------------------------+
$ agent-traces ls --all --json | wc -l # every agent session on this machine
2998
$ gh repo list royalaid --limit 100 # every repo that moved since april
51 touched, 12 public, 39 private
$ cat summary.txt
tl;dr
I run several coding agents at once (Claude Code, Codex, OpenCode, Grok,
a resident agent called Hermes) on a few machines. Most of what I built
since May is the plumbing that lets them hand work to each other, remember
what happened, stay safe unattended, and ship results somewhere I can read.
Nothing here was planned as one system. Each piece answered a failure of
the piece before it. This page is the map after the fact.
pulse
commits / month (personal repos, de-duplicated, work excluded)
may ######## 154
jun ##### 109
jul ########################## 527
aug ########### 220
sep ####################### 459
oct ############################################# 891 (10 days)
sessions indexed 2998 (about 2400 once empty probe sessions are dropped)
harnesses claude, codex, t3, opencode, grok
the seed (may)
2026-05-01 a "second opinion" suite
Claude drafts. Codex independently checks the plan, the doc,
the root cause, the fix. Two model families, one verdict.
second-opinions ---> delegation skills ---> roles, not models
| |
+--> goal-forge ---> unattended runs +--> handoff mailbox
That one habit (never trust a single agent's word) turned into the
orchestrator / worker / reviewer split that everything later uses.
Same month: a first keep-it-alive fix so an agent runtime survives
a closed terminal. Boring, and it turned into half the infrastructure.
timeline
may second opinions . agent runtime kept alive . first skills repo
jun 06-11 "commander": one thread steers three repos on goal loops
06-11 lore born (agent work-trail memory)
06-19 code-search MCPs wired into every harness
06-29 one-command publisher + 3D blog homepage
jul delegation skills . T3 Code fork . first destructive-command guard
aug Hermes fork + nightly lane sync . slop (artifact host) . usage tracker
sep 09-04 harbor goal written . 09-26 telepathica . 09-28 agent-traces
oct 10-03 Rust guard + patched Codex sandbox . 10-05 harbor absorbs the small repos
how it weaves
YOU
threads . chat . links . approvals
|
+--------------------+--------------------+
| | |
HAND OFF REMEMBER PUBLISH
telepathica <--> lore + chartroom <--> slop + royalaid-sites
(mailbox) (notes, docs) (pages, one link)
| ^
v |
WORKERS ---- trace ----> agent-traces
claude . codex . opencode . hermes
|
+--- SKILLS ----- one source, five harnesses
| +--- agent-guard (deny the irreversible)
| +--- bosun (tool friction -> the repo that owns it)
|
HARBOR ---- backup . catalog . health . restore drills
|
own-tools-sync ---- every merged change, every machine, ~15 min
four loops
[1] delivery
task -> orchestrator makes the design calls -> closed contract
-> worker in its own worktree -> reply "done" or "blocked"
-> independent review -> merge -> sync installs it everywhere
[2] memory
session -> traced -> lore distills overnight -> notes I review
-> injected into the next session's start
[3] friction
agent hits a broken tool -> bosun report, keeps working
-> filed in chartroom -> fix lands in the owning repo -> loop [1]
[4] publish
anything visual -> one HTML file -> slop -> one stable link
(you are looking at the output of this loop)
the parts
core
harbor monorepo: backups, service catalog, health, a folder per tool
skills ~45 hand-written skills, one source, read by 5 harnesses
telepathica mailbox for handing work between agent sessions
agent-traces cross-harness session search and handoff public
lore memory: captures agent work, distills it, feeds it back
chartroom plans, specs and handoffs kept outside git
bosun routes tool friction to whoever owns the fix
agent-guard deny-only safety hook for unattended agents (Rust)
own-tools-sync keeps every machine on the latest of all of the above
runtime
hermes resident agent with chat, cron and a task board fork
hermes lane sync nightly rebase of my fork onto a fast-moving upstream
t3code thread GUI that shows all the agents fork
codex patched so unattended runs stop hanging on Windows fork
helium-mcp one shared logged-in browser for every harness
publishing
slop agent-oriented static artifact host
royalaid-sites one command, local folder to live subdomain
dotfiles shared agent instructions, identical everywhere public
side quests
homepage blog rendered as a 3D illusion, then a playable castle
fc-vid-uploader Farcaster client tooling: X video into casts, feed work
rift-wizard-explorer offline game build planner
leet-code-fixer study app with an agent coach (ancestor of lore)
clipsync clipboard sync across phone, Mac, Windows
novaconnect tray app for headset power state and audio switching
fxtwitter-copy-link tiny Chrome extension public
patterns that kept repeating
1 every failure grew a layer
a broken backup became a fleet repo, then an unnoticed outage became
independent monitoring. A hung agent became a patched sandbox.
2 roles, not model versions
orchestrator, executor, reviewer, light. Models change monthly;
skills never name one.
3 receipts over claims
"fixed" has to come with a check someone else can rerun.
4 one source, many hosts
edit once, a 15 minute loop carries it to every machine.
5 carry forks, don't hand-merge
patch sets re-applied by scripts and by models that resolve their
own merge conflicts.
6 consolidate late
many small repos in summer, one monorepo in october, old repos left
behind as one-way mirrors.
7 autonomy needs a floor
agents run freely; a small guard blocks the few things that cannot
be undone.
method
Mined with agent-traces (sessions since 2026-05-01, the first day the local
stores have anything) and gh (repos pushed since 2026-04-01), fanned out
across parallel read-only scouts, one per area. Counts are approximate.
Work projects and archived experiments are left out on purpose.