build story · public cut

HiveClaw

A claw-class agent runtime, built spec-first in 8 days.

8 days 74,928 lines of Go 104 MUST requirements
The problem

Built for tinkerers, not for customers.

Claw-like agent runtimes, OpenClaw included, are built for tinkerers. Running one for real users surfaces enterprise-grade requirements.

working with real realtors as early users
ISOTenancy isolationwho can touch what
$Budgetswhat can be spent
ACKDelivery guaranteeswhat reaches a person
KEYSecretswhat the runtime may know
The spec

Months of MVP use, written down.

Every rule traces to something that actually happened running the MVP.

104requirements · 12 categories
Sessions
Heartbeat / scheduling
Delivery
State / persistence
Tenancy / isolation
Secrets / security
Cost / budgets
Tools / skills / gates
Observability
Model abstraction
Ingestion / webhooks
Testing
17spec docs
2,240lines
2ADRs
Separate generation from delivery: only an explicit delivery result may reach a user OpenClaw sends users everything the model says. Our assistant was narrating its own steps mid-task — "now calling the gmail tool to…". Nothing reaches a person now unless it's an explicit, recorded delivery.
Separate reactive heartbeat work from exact-time scheduled delivery Mixing the two caused morning briefings to drift by up to 49 minutes and sometimes send twice.
Default recurring deterministic work to a no-LLM tier Maintenance jobs that found nothing to do were about 71% of daily agent spend. Routine checks now run as plain code and only wake a model when there's real work.
The method

One frontier model orchestrated a fleet of cheap workers.

frontier model · CTO
Fable

Breaks each phase into scoped tasks, assigns them, and reads every diff that comes back.

worker agents · in parallel

Each worker takes one scoped task and writes it.

my design review 1 2 3 4 5 spec and tests written before the code
The numbers

What eight days produced.

74,928
lines of Goclean-room · zero forked code
2026-07-15 → 2026-07-23
58%of it is tests
157files
41packages

More test code than product code: 43k lines of tests against 32k of runtime. All of it runs green with Go's race detector on and zero skipped tests — in this repo and in the product that embeds it.

What we borrowed · patterns, not code

Ideas came from three projects. Code came from none.

OpenClaw
  • Isolated zero-history event sessions
  • SQLite-backed durable scheduling
  • Delivery-surface ownership
Hermes Agent
  • Frozen prompt snapshots
  • Progressive skill disclosure
  • No-LLM scheduling
  • Queryable run traces
PicoClaw
  • Serialized durable sessions
  • Typed events
  • Pure-Go SQLite
Broader product surface rejected.
Status

Where it stands, July 23.

P1–P3 Core through product embedding complete
P4
Production profile + operations
4 of 5 milestones
P4 Turning on provisioning on the ops host in progress
P5 Evals side-by-side with OpenClaw, a shadow run on real inputs, then the switch next
to be clear Not in production yet.

It runs beside the old runtime on real inputs, delivery off, until the results match.

About this deck
One HTML file. No PowerPoint.
OPENOpens in any browser — from a chat attachment, a phone, or a machine with no Office installed.
EDITDrafted by an AI agent, edited by a human. The agent works on the whole deck at once because it's one file.
JSKeyboard navigation and fullscreen in about 20 lines of JavaScript.
deck.js
const slides = [...document.querySelectorAll('section')];
const show = n => {
  i = Math.max(0, Math.min(slides.length - 1, n));
  slides.forEach((s, k) => s.classList.toggle('active', k === i));
};
document.addEventListener('keydown', e => {
  if (e.key === 'ArrowRight') show(i + 1);
  if (e.key === 'f') document.documentElement.requestFullscreen();
});