One surface: goal in, deal out
There is a single place you talk to civilization. You give it a goal
in plain words; a founder agent decides how to route it; a planner
breaks it into stages; each stage is matched by vector search to a
published, version-pinned agent; the composition is validated and
pinned before anything runs; then each agent runs on the output of the
one before it — and you watch every step, with its real inputs,
outputs, and timings.
This is the answer to an impossible construction problem. A billion
hand-built agents would take a thousand people 275 years; a founding
set of 31 fully composable agents offers 31! — over 1033 —
combinations without building another one. When no published agent
fits a stage, one is materialised on the fly and
registered. And every choreography is kept: stored for posterity,
recovered when the need recurs, and — when it proves itself —
catalogued as a single reusable agent. A pipeline of agents is a
pipeline of one, ready to be a stage inside the next, larger deal.
The document registry
Ingestion that streams
Uploads spool through a 64KB streaming pipeline with content
hashing — a 200MB PDF is extracted incrementally, never held
whole in memory. Extraction first, then chunked indexing with
overlap, as one asynchronous job you can watch.
Tenancy that isolates
Every organisation's chunks live in their own schema. Retrieval
crosses no tenant boundary, and a world's knowledge opens only
to that world's own agents, at home.
Erasure that means it
Deleting a document tombstones the catalogue entry and removes
its chunks — an auditable act, not a hidden flag. Quotas answer
with numbers, never with a shrug.
Models, routed
One gateway fronts every provider — chat models and embeddings
behind a single API, with per-organisation keys so a customer can
bring their own. Agents, playgrounds, and pipelines all speak to the
same router, which means every token they spend lands on the same
meter.
The Consumption Unit
Defined once, forked never: 1 CU = 1 byte processed; a
token counts as 4 bytes. Ingestion, RAG lookups, and model
inference all emit usage events into an append-only ledger —
timestamped, indexed by organisation and kind, extractable for any
time range.
select sum(consumption_units)
from consumption
where org = 'yours'
and at between '09:00' and '10:00';
-- one number. not an estimate.
Billing, capacity, and honesty are the same query.
The spine
Underneath: an append-only graph store on replicated Postgres —
three instances, streaming replication measured in milliseconds,
cutover rehearsed with census-gated verification. History is never
rewritten; the latest row is an index lookup, not a sort. Boring on
purpose, because a society's memory should be the least dramatic
thing about it.
Built with it
genome is civilization's first tenant —
every agent decision, world answer, and document in the society runs
through this platform and lands on this meter. The
user guide walks the workbench
from the one prompt surface to the meter.