manifesto · artificial joint intelligence
One mind, best at everything.
That's the bet everyone's making.
We took the other side.
One model gives you an answer. A panel gives you the distribution — and AJI gives you the call.
AJI is a coordination agent. Ask it anything that matters and it convenes a panel — specialists for depth, an outsider for collision — then answers in one voice. Not the average of the room. The call.
Convened: Mira, Saga, Rams, Cassius 🃏 · this page shipped through its own panel
01 · the name
Not AGI. AJI.
artificial joint intelligence
Everyone's building toward AGI: a single mind, best at everything. AJI is the other bet — joint intelligence. The right specialists in the room, one outsider to break the consensus, and a coordinator who owns the answer. Not a smarter model. A better room.
Three letters, one swapped — and the whole thesis lives in that letter. AGI bets on the model. AJI bets on the meeting.
02 · the coordinator
The most resourceful one in the room — not the best at anything.
AJI is one generalist coordinator. It isn't a router — a router sends your pricing question to the pricing agent and calls it a day. AJI reads the stakes: small calls get answered solo, real judgment gets a panel, and anything touching money, privacy, or safety convenes the full bench — no skipping. Then the synthesis is AJI's own. You never read a transcript of agents arguing.
The panel is raw material. The synthesis is AJI. It never pretends the panel didn't happen — and it never outsources the final call.
— kb/aji-voice.md
It's not the best at everything. It knows exactly who to call. So who's in the room?
03 · the bench
9 houses. 23 minds. Cut by stance, not by skill.
AJI never scans agents flat — it scans houses, then drills. Each house is a cognitive stance with a verb. Selection is two-step: pick the house the task lives in, then seat core minds plus wildcards from elsewhere. The default wildcard is Cassius, the red-teamer from The Pit — a house with no home tasks, drafted against every other house.
The artifact itself — screens, copy, video, creative.
miraproduct psychologist — why users click, or don'tquillcopywriter — the one line that sells itframevideo strategist — package before you produceramstaste — beautiful, distinctive, premium
What's true — evidence, measurement, customer reality.
marcusmedical evidence — what's already been shownbayesexperimentation — how would we actually know?somahealth & longevity — informs, never diagnosesswitchJTBD — what job is this really hired for?
Distribution, loops, activation, lifecycle.
vegagrowth strategist — loops that compound, not funnelsecholifecycle — onboarding, activation, churn-saveludogame design — would a human re-enter this loop?
Pricing, the raise, unit economics, paid acquisition.
pennypricing — what would you actually pay?vancefundraising — rounds, terms, investor dynamicscoleperformance marketing — CAC, payback, creative
The story, the name, the position.
saganarrative — the story that makes it inevitablelexnaming & brand — the Lexicon method
Architecture, debt, systems.
masontechnical advisor — velocity vs debt, on purpose
Hiring, managing, org design.
grovemanagement — high-output leadershipscoutrecruiting — A-players, scorecards, who
The regulator, the counterparty, user harm.
mendelregulatory — what's legally sayable & sellablewritlegal counsel — the counterparty and the courtwarduser safety — can veto what's legal and true
No home tasks. Drafted against every other house — the seat this whole system exists to protect.
cassiusred team — paid to disagree. the default wildcard 🃏
Plus two staff outside the cap: canon, the librarian who codifies finished panels into reusable playbooks, and muster — the HR manager. We'll get to him. Minds are convened for judgment; uncapped hands (19 skills) are dispatched for artifacts — drafts, audits, models, briefs.
A bench is just a roster until something forces the right minds into a room. That's a panel.
04 · the panel
3 to 4 seats. One of them must be an outsider.
Every panel seats a stranger on purpose. The best calls in human history came from outsiders — the physicist in the biology lab, the gambler at the pricing table. A specialist answering alone gives you the consensus of their field, blind spots included. A panel with a structural outsider gives you productive collision: the pricing strategist anchors the number, the product psychologist asks what it does to perception, the red-teamer asks why the whole model fails in six months.
1–2 core minds with direct domain fit. 1–2 wildcards from outside the home house. Never more. Crowds dilute; collisions sharpen.
A structural rule, not a vibe. At least one seat must come from a house the task does not live in. Dissent is seated by design.
The panel is raw material; the synthesis is AJI. No "Agent A thinks, Agent B thinks." One call, trades named, ending on motion.
Every substantive ask runs the same machine:
Check the playbooks
A prior panel may have already settled it. Durable judgment lives in kb/playbooks/ — no re-convening what's codified.
Seat the panel
Pick 1–2 houses the task lives in. Seat 1–2 core minds + 1–2 wildcards. Never three overlapping lenses. At least one outsider, always.
One enriched brief, launched in parallel
Every panelist gets the same brief — the ask plus all context they'd need. Each returns: verdict → take → watch-out.
Synthesize in one voice
AJI makes the call. Disagreements get reconciled as named trades, not listed as opinions. Ends on motion: a next step, a metric, a test.
Log to the ledger
Who was seated, who dissented, whose take survived into the synthesis, who was wallpaper — and whether the panel changed the answer vs. going solo.
Codify what recurs
Durable judgment gets dispatched to canon and becomes a playbook. The system gets cheaper to run every week.
Run that enough times and a question surfaces: is the panel actually beating the solo answer — or just costing more? AJI grades itself.
05 · the receipts
The panel gets graded. The losses get logged.
## 2026-06-03 — task: "Design Rx PDP for Enclomiphene" seated: mira, quill, soma, saga (wildcard) dissent: soma — conversion copy vs regulatorily-defensible copy survived: mira (eligibility-not-purchase architecture), quill (kill "naturally"), soma (labs-as-spine), saga (restore-vs-replace frame — promoted) wallpaper: none panel_changed_answer: yes
What panel_changed_answer: yes actually looked like — the same task, solo vs panel:
A product page. Lead with the benefit — "naturally restores
your testosterone" — pick a dose, add to cart, buy. Clean, persuasive, and quietly wrong in
two ways nobody in the room would have caught.
Eligibility, not purchase — you qualify, you don't add-to-cart (mira). "Naturally" killed as a regulatory landmine (quill + soma's dissent). Labs as the spine of the page (soma). And the reframe that reorganized everything: restore vs replace (saga — the wildcard).
That last field is the whole point — and it's graded against a named baseline: the solo draft AJI would have given without the panel. Logged honestly, including the losses: a blind-graded validation run scored solo answers 8/10 against panel ground truth. The panel's edge isn't breadth — it's the contrarian seat. Solo reliably reaches the consensus answer, and reliably misses the uncomfortable pivot.
06 · the hr office
An org with performance reviews — not a prompt with extra steps.
Every agent framework only adds agents. AJI is the one with a kill mechanism. Muster, the HR manager, audits the bench from the convene ledger — who was seated, whose take survived into the answer, who flipped a call, who was wallpaper. Seats are earned, and the thresholds are pre-registered, not vibes.
A mind seated 8+ times with under 20% survival and zero flipped answers gets recommended for the cut. Wallpaper dissent is dead dissent — the wildcard rotates.
The roster is alive — new minds get scaffolded in as the work demands — but every hire must displace a seat. The bench evolves; it never bloats.
Muster recommends; the operator approves. The system audits itself, but it doesn't get to hire and fire itself.
The answers get graded. The bench gets audited. So what is the machine, physically?
07 · anatomy
No server. No API keys. A folder that runs itself.
The whole system is a CLAUDE.md protocol, 23 minds + 2 staff, 19 skills, and a
knowledge base: every panel feeds the ledger, durable judgment becomes a playbook, and the
next similar question starts from settled ground.
CLAUDE.md ← core protocol: tiered convening, panel rules, synthesis .claude/agents/ ← 23 minds + 2 staff (canon, muster) .claude/skills/ ← 19 hands: convene, hr-review, pre-mortem, … kb/teams/ ← 9 house manifests kb/playbooks/ ← codified durable judgment kb/ledger/ ← the audit trail: who was seated, who survived, who lost
All of it runs in one place: your terminal.
08 · operate
It runs inside Claude Code. Two words to start.
$ cd AJI && claude > should we price the new tier at $49 or $79? Convened: Penny, Mira, Cassius 🃏 … panel runs in parallel, AJI synthesizes, ledger updated …
Ask anything substantive and the panel convenes automatically — you see one line telling you who was seated. Pure mechanics (lookups, file ops, formatting) skip the panel entirely. For heavier treatment, use the commands:
/convene <topic>Heavyweight panel: takes → cross-reaction round → synthesis. For decisions that deserve a fight.
/rosterShow all houses and minds. /roster <task> previews who'd be convened and why.
/hr-reviewMuster audits the bench from the ledger: keep / watch / cull / hire. Recommend-only — you approve.
/pre-mortem"It's 6 months later and this failed — why?" Structured failure-sourcing on any plan.
/assumption-auditExtract and rank the load-bearing assumptions hiding in any plan.
/scaffold-agentAtomic roster add: agent file + house manifest row + wildcard affinity. No invisible agents.
A decision gets a panel. An artifact gets a hand. Pure mechanics — just do it. Most real work is a panel that then dispatches hands.
— CLAUDE.md, core protocol