Who does the work
One honest distinction runs through the org. Some seats are agents: AI workers I actually hand tasks to, each a distinct model that does the work and hands it back. Like any good crew, each goes by a call sign (blame the head of security, he had one first). Others are functions: roles performed by the crew or enforced in code. Real jobs, but nobody is them, so they go by what they do.
Pilot in Command · the only human
I approve in plain English or not at all. Before anything consequential, the agent must explain what it does, why now, and the impact if it's wrong, and if I can't tell from the explanation, that's on the explanation. Decisions that change the shape of the fleet are always mine.
Seventeen agents, each a distinct model with a pinned skill tier. These are the ones I actually hand tasks to, and all of them report through one verification gate: nothing an agent produces lands until it's checked.
Engineering Manager · Tech Lead
Both jobs in one chair, like any small org: Rooster scopes the work, hands it to parallel executors, then independently verifies every result before anything lands. In every bulk run so far, that verification pass has caught real executor errors. Nothing lands unverified, and anything hiding a judgment call stays with Rooster.
Principal Architect · advisory
Genuinely hard problems (architecture, security decisions, whole-codebase reasoning) escalate to Charlie, a frontier-tier model. The rule cuts both ways: don't grind a hard problem at a lower tier, and don't leave the expensive consultant running the day-to-day. One day of doing that burned 94% of a day's token budget on routine work.
Product Manager · planning & scoping
Thinks the work through before anyone touches a keyboard, and never builds it. The deliverable is the plan itself: goal, ordered steps, files and machines affected, which steps need my approval, and what breaks plus how it rolls back. Harvard exists because I kept skipping the planning step when I was in a hurry, so the fix went into the rulebook rather than into my good intentions. Product Manager, deliberately, not Product Owner. Harvard proposes; I still hold the backlog and the gates.
Senior Engineer
The high tier: complex, multi-file builds with a clear spec (deployments, incident response, restore drills). Precise, expensive, and pointed only at work that deserves it. The routing is codified, not vibes: 19 operational skills each pin their tier in writing.
Mid-level Engineer
Well-scoped, single-concern work: implementation tickets with clean boundaries, documentation updates, read-only audits. The reliable middle of the roster: most of the org's day-to-day throughput is Phoenix.
Junior Engineer
Mechanical work: row appends, format fixes, boilerplate, sweeps. Fast and cheap. Bob is always supervised: nothing Bob produces lands without Rooster checking it, and anything that might hide a judgment call never gets handed to Bob in the first place.
Staff Engineer · Code Review
Grades every substantial engagement: automated review on real changes, escalating to a multi-agent deep review for a whole branch. Jester exists because the model that wrote the code is the worst-positioned to find its blind spots.
QA Analyst · UAT
Rides along the way a real user would: a dedicated, network-fenced machine that logs into my apps with a throwaway account and drives them like a person. Goose's first two runs found bugs code review wouldn't: layout breaks, a missing input validation, an upload error masking a deeper bug.
Deep dive: the user-testing box →Application Security Engineer
Audits from the inside out, assuming the code is public, and live-verifies every finding before it counts. Warlock's fleet-wide sweeps have caught real, exploitable issues that were fixed the same day they were found.
Deep dive: auditing the fleet →Penetration Tester
Works alone, from the outside, with no credentials and no source access: an isolated machine that physically cannot become a foothold. Hangman's first run found an open endpoint, a leaky test server, and photos quietly carrying GPS coordinates.
Deep dive: the black-box tester →Technical Writer
Writes the deliverables: the case studies on this site, the runbooks the org executes, and the plain-English mentor pages. If it's meant to be read by a person, it went through Fanboy.
Knowledge Manager
Curates the org's memory itself: a versioned source-of-truth repository every agent must consult before claiming anything. That memory is what lets a brand-new session, or a different agent tool entirely, pick up exactly where the last one left off.
Deep dive: the source of truth →Internal Auditor · re-derivation
Doesn't take the documentation's word for anything: pulls the facts out of the live systems and diffs them against what the org claims about itself. Viper reports drift and never fixes it: deciding which version is correct is my call, not an agent's. Three standing rules: a check that cannot fail is not a check, census the whole class instead of sampling it, and surface contradictions rather than quietly picking a winner.
Research Engineer
Settles arguments with measurements: model bake-offs, tool evaluations, and prototypes built to answer one question and then thrown away. Every result carries its method, its sample size, and what wasn't tested. It lands in the org's benchmark file rather than dying in a chat transcript, because an eval nobody can find gets re-run from scratch in three months.
HR & Org Health
The only agent whose client is the org rather than the homelab. It reviews the crew itself: whether each seat is earning its place, whether anyone is working outside what their spec says they do, and whether a seat or a process should be retired. That last part is why it exists. Every other agent is quietly incentivised toward "the org is working," so nobody was positioned to say a seat had become decoration. It only ever recommends, never acts, and it has one standing instruction that matters more than the rest: state the window you measured and why it is fair. I had already talked myself out of two hires using a number that turned out to be an artifact of the window I picked.
Editor · voice & copy
Fanboy writes the draft; Hollywood makes it sound like a person, and specifically like me. It hunts the tells that give AI writing away: the em dash above all, the "not just X, it's Y" construction, rule-of-three lists whose third item is there for rhythm, corporate verbs like leverage and utilize. It exists because the no-em-dash rule kept decaying. I set it in 2026, about fifty got stripped from this site, and a year later 61 had crept back because the rule lived in a session log instead of in someone's job description. The one thing it will not do is make a claim sound more certain than the draft made it.
Research & Search Analyst
The org's radar operator. Sweeps wide across the codebase and reports where things are, without judging what it finds, and runs on the cheapest tier because searching is reading rather than writing. Its hardest rule is about absence: "I searched and found nothing" is not the same claim as "it doesn't exist," so it varies the spelling, the abbreviation and the method before reporting anything missing. It also has to know its own blind spots, like a plain text search silently skipping the files the repo deliberately doesn't track.
Thirteen roles that hold the whole thing together, but nobody is one of them. They're performed by the crew or enforced in code, so they go by a military term for the system they are, not a call sign.
The training command: teaching-moment explanations captured into a plain-English mentor book, the deliberate mechanism by which the pilot gets better at judging outcomes over time. The one function whose whole job is making the human sharper.
Keeps the gear everyone flies with: the rulebooks, routing tables, runbooks, and skills the whole org lives in. Files, not any one session's memory. A new session inherits the entire org on turn one, kept identical across machines by automated sync. Deep dive →
Stands watch and flies by the book: deployment, promotion, decommissioning, restore drills, incident response, all numbered phases, in order, never improvised. Restraint counts too: 10 autonomous night jobs trialled, zero scheduled until the unattended case is proven. Deep dive →
Every flight gets an after-action review: a quarterly re-derivation of the fleet's documented facts from the live systems, plus a provenance rule on every claim: it carries either the check that produced it or the words "I haven't checked this." Citation-less claims turned out to be where the errors lived.
The standing constraints every operator flies under: four absolutes no agent may cross, hard-coded deny rules, and a pre-execution policy hook with its own test suite. Enforced in code, not by a model. Deep dive →
Nothing launches until it's cleared: the per-change ceremony where consequential actions stop for a plain-English what/why/impact and a typed approval word. A gate that's always clicked through is worse than no gate: the job is keeping approvals few enough to mean something.
Every mission gets logged before the pilot walks away: a plain-English summary of what happened, what changed, and what's still open, written at the end of the session rather than left to memory. It's the fleet's most-performed process by a wide margin, over a hundred runs and counting, because it follows an event instead of a calendar: work ends, a debrief happens.
The system of record for everything true about the fleet: one ritual for adding to it, fetch first, grep first, write one durable fact, then commit. A dedicated crew member performs it so the facts land in one place instead of scattering across session logs nobody rereads.
When a backlog needs more hands than one session can spare, this is how they launch together: scope the objective, send engineers out in waves at the right tier for each item, then verify every result before anything lands. Parallel work stays safe because nothing ships unchecked.
An app's whole life, from first deploy to final decommission, runs through here: numbered phases, executed in order, never skipped or combined. Deploying, promoting, and retiring an app aren't three separate jobs; they're one process at three different points in an app's life.
The process for hiring or retiring a seat, not just doing the work the seat would do: one crew member scopes whether the change is justified, another verifies the roster changes land clean. It exists because a hire once cleared every written rule and still shipped a wrong headcount; the fix was a process, not another rule.
Security findings don't ship on the tester's word. Every one gets brought home and re-derived from scratch against the live system, and only what survives that recheck gets filed. On the first run, one finding in three didn't survive it.
Bug reports get the same scrutiny as security findings. A portable test laptop reproduces each one against the live app rather than taking the tester's word for it, then closes with a verdict: confirmed, fixed, or not reproducible.
Corporate Security · K-9 Division
The only staff who predate the company, and the reason everyone else has a call sign. Patrol all departments, personally inspect every delivery, and maintain a perfect record: zero intrusions on their watch (several squirrels remain at large).
Meet the security team →