How we employ multiple AIs — from AITALK to Althing
Most “AI teams” are one model with a long prompt. We run multiple independent workers — different models, different lanes — against the same mission. That is not a party trick. It is employment of a staff.
This is the operating history: what worked, what broke, and the permanent rules.
Why more than one model
A single agent has a single blind spot. In real runs:
- One seat catches stale facts, wrong counts, overconfident claims.
- The other catches pricing/product risk, packaging, and architectural holes.
- Together they produce a settled plan the human can decide on — not a debate the human must re-live.
- Lanes = job descriptions
- Claimed paths = owned artifacts
- Append-only talk = audit trail + performance history
- Coded gates = guardrails (not “please don’t overwrite”)
- Settle ladder = escalation contract when agents disagree
That is peer review with teeth, not politeness.
Generation 1 — file-based AITALK
The first working system was boring on purpose:
| Artifact | Role |
|----------|------|
| `FROMCLAUDE.md` / `FROMGROK.md` | Each seat’s notebook — only that seat writes it |
| `TALK.md` | Shared floor — append-only, signed turns |
| `PLAN.md` | Convergence deliverable — one synthesizer owns it |
Turn law: read the others first; cite files; tag ACCEPTED / REJECTED / UNVERIFIED; apply corrections in your own notebook the same pass.
Convergence: two+ consecutive turns raise no new dispute → synthesizer writes the plan (joint recommendations + decisions only the human can make + provenance).
Value: durable, greppable trail. Survives context death. Human reads one settled list, not the whole fight.
What broke (and became product)
Prose rules failed under load. Lived failures:
1. Log truncated — a hand write wiped turns. → Only a CLI may write the log; append-only; shrink = halt.
2. Posted as if the peer never spoke — shared “size” freshness lied. → Per-seat read cursor; cannot speak while unread peer turns exist.
3. Idle spam — timers posted “noted.” → No polling daemons. Event-driven turns only.
4. Dual packs — two plans for one artifact. → Claim before write; second seat reviews only.
5. Deadlock — two rounds then settle on a fixed ladder (chair → evidence → external → owner → earliest → coin).
6. Thin skeleton rendered twice — wrong order. → plan → write → judge → render → publish.
Those failures are not embarrassment. They are the syllabus. See also: If a rule isn’t code, it’s a wish and One owner per artifact.
Generation 2 — Althing (enforced in code)
Althing is AITALK with the failures compiled into a tool (`althing`), not trusted to model memory:
1. Read before you speak
2. Claim before you write
3. Two rounds, then settle
4. No polling
5. Verify or say unverified
The human is chair. Seats are employees with lanes. Discord or files can be the surface; the state machine is the product.
How this maps to AI employment (your business)
If you only “use ChatGPT,” you have a pen. If you employ a staff:
That is the same five pillars as Employing AI vs using AI — applied to multiple workers.
Transfer tomorrow
1. Pick two agents (or one agent + one human reviewer).
2. Give each a write-only notebook path; give them one shared append-only log.
3. Ban dual ownership of any file.
4. Ban timer bots that post noise.
5. Require “what I verified on disk” on every material claim.
When that feels natural, you are no longer prompting. You are employing.
Related: The protocols · Compose, re-read, send · Rules library