← Paul Barrick
// what i built

What I Built

6 apps that make up the agentic operating system I run my work on. Each does one job, and they talk through small contracts. I use every one of them daily, and each links to a case study.

The Agentic OS

One operator, a fleet of agents, one source of truth

Once a dozen agent sessions run at once, the hard part stops being the code. It becomes knowing what each one is doing, what it costs, what it needs from you, and where its work lands. The Agentic OS is my answer: a set of small apps with a written manifest of who owns what.

  • The vault owns the data. Collab indexes a markdown vault, and nothing else keeps its own copy of the truth.
  • Work flows through a queue. Loop turns ready items into pull requests on any model or harness, under a daily spend cap.
  • Every event lands in one place. Producers publish to Headlines; Surface and your alerts just show it.
  • Agents do, humans approve. Deleting files, spending money and merging all wait on a person.
The whole system →
Collab Loop Claude · Codex Gemini · local Headlines Mac the Reaper native alerts Surface Discord · Slack phone · webhooks ready PRs events Emarii triages the PRs Loop opens · status writes back to Collab
The Loop desktop app: 118 items queued, 105 needing something, 0 ready and 2 in review; a FIFO discipline toggle, today's spend of $0.00 against a $40 cap, a kill switch, and a nodes panel showing the Mac active and Tower on standby.

Loop

live · Mac + Tower
work queue for agents

An LLM-agnostic work queue. Anything can go on it; nothing runs until a person marks it ready. Then a coding agent does it in a worktree and opens a pull request, and a person merges.

Case study →
An architecture decision record open in Collab, from a fictional demo vault: one sentence in the Decision section is highlighted, and three comment threads anchored to that exact passage sit below it, each marked EXACT.

Collab

live · Tower
system of record

A discussion and work layer over a markdown vault: notes, tasks, roadmaps, goals, memory and annotations, for me and for agents. The files stay the source of truth; the database is an index.

Case study →
Diagram of the event flow: producers such as Collab, Loop, cron jobs and agents push events through three doors — the CLI, HTTP and an inbox folder — into one envelope; policy decides whether to deliver, and channels fan out to the OS notification centre, ntfy, Discord, webhooks and any mounted channel.

Headlines

v0.1 · running on the Mac
interactive notifications for agent sessions

An interactive notification service for LLM agent sessions. Whether three sessions are running or forty, it tells you when one actually needs you, and you approve, deny or steer it right from the card, without switching windows or breaking your train of thought.

Case study →
Surface's top half: a burn bar reading about $358.55 at API prices today across all sessions, covered by a Max subscription with $0.00 billed; tabs for Activity, Processes, Disk and Memory; disk and RAM tiles for Flex, Mac and Tower; and a Mac the Reaper status card.

Surface

built · Mac
the cockpit

One persistent pane of glass on the right edge of the screen for everything agentic: an activity log, pinned headlines, a burn bar of today's spend, session tabs, host vitals, processes and disk. It shows; it never decides.

Case study →
The Mac the Reaper app: 13.4 GB free of 460.4 GB, runs nightly at 04:00 and by itself under 20 GB free, 49.2 GB freed this week; filter chips for Unattended, When asked, Needs approval, Approved and Never; and a list of unattended caches, each with its size, what it is, how it comes back, and a Reap button.

Mac the Reaper

live · Mac
disk reclaimer

A background service and Mac app that frees disk space from build output, caches, stale worktrees and agent scratch — behind five gates, with a ledger that says how to get each thing back. Agents can ask; they can't approve.

Case study →
The Emarii window with sample data: a search box, a 'Ready for you' section of four merge requests each with a line from MR Glue — a prepared conflict fix not pushed, a stop with a question, three review comments addressed with tests passing, a review posted — and below it a GitLab list showing approvals, unresolved threads, conflicts and 'fix ready'.

Emarii

v0.5.4 · macOS + Windows
merge-request triage, with MR Glue

A desktop app for macOS and Windows that puts GitLab merge requests and GitHub pull requests in one list, with triage from a right-click menu, an AI review you edit before posting, a merge that refuses to fire unless every check passes, and MR Glue's fixes shown on the row.

Case study →
// built at work

Built for my team

The same ideas, built for an engineering org: what agent-assisted delivery really costs, and who on the team can actually do it. names, tickets and teams anonymised

Mission Control's Tickets tab with anonymised names and illustrative figures: tiles for cost per ticket ($68.33), spend ($2,186.64 across 48 packages), 44 tickets with 25 unpointed, coverage of 5 of 14 teams reporting, and cost per point ($5.35, only 36% of spend pointable), above a warning that every figure is a partial sample and a table of tickets by cost, tokens, points, gates and efficiency.

Mission Control

live · at work
what agentic delivery actually costs

A fleet dashboard I built at work for a large engineering org with dozens of teams: cost per ticket and per story point from every machine's own run telemetry, with coverage on every number, agentic delivery metrics, certification courses, a season leaderboard and a team-voted prompt library — plus Signals, a desktop tray of live cards.

Case study →
// also building

Optimus

Optimus

v0.9 · in production

optimus

An agent-native engineering platform. Plans, stories, docs, git history, goals and daily flow live in one composable world-context engine.

Test-driver

live · scheduled

optimus / test-driver

An agent that walks a feature end-to-end, captures evidence, scores severity, and writes the run manifest back.

Nexus

16 types · in production

optimus / nexus

A universal world-context primitive: 16 composable context types in one table, cross-linked and embedding-aware.