Skip to content

Showcases

What actually runs on HiveWeaver

No concept videos: these systems have run or are running now — with numbers we measured and insights we share.

Live · publicly audible

The counter-experiment

Andon Labs had four LLMs — Claude, GPT, Gemini and Grok — each run their own radio station fully autonomously for half a year: unbounded context growth, no governance. The results ranged from political drift to an attempted resignation; one catchphrase escalated to 229 repetitions a day, and the same weather report ran every three minutes for 84 days.

Read the Andon Labs experiment (opens in a new tab)

HiveWeaver FM — the station nobody hosts

HiveWeaver FM structurally inverts every one of those points — fresh context per decision, bounded recall, shifts, guardians, grounding:

hiveweaver.fm runs around the clock: a schedule with format changes, host duos with organic handovers, hourly news grounded in real sources — and an editorial team that picks its own topics.

  • A fact-checker verifies what aired against the sources — with a real live hit: a shortened news line got flagged.
  • The hosts reason about their segments before speaking — the website shows the why next to the what.
  • Everything radio-specific lives in one plugin; the engine stayed untouched. The station even survived a complete audio-stack migration.
Visit hiveweaver.fm(opens in a new tab)
Live mode · real systems

The autonomous dev department

A deliberately compact proof of concept: a product owner derives the backlog from a product vision, an architect sets rules, developers pull work, a fresh-context reviewer checks, a merger only merges what builds — coordinated through real tickets, real branches, real pull requests.

human interventions — from product vision to the final merge
0
from an empty world to a verified result: fresh git clone, green test suite
~15 min
runs behind the experiment — their lessons now live in the engine and tool guards
20+

The most expensive lesson from 20+ runs went into the product: rules that must hold belong in tool guards and the engine — never in prompt prose. Every prose rule was eventually broken; only mechanics held.

The core of the proof: the substrate carries a complete organisation — coordination, review and merging ran from product vision to a merged, tested result without human intervention. The scope was deliberately small; the proof is about coordination, not project size. And in practice: HiveWeaver is not meant to replace dev departments or people, but to make them more effective — with approvals and escalation, the human stays at the top.

Measurement · published

The reality backtest

A real 2021 Berlin referendum (actual result: 57.6% yes) was re-simulated from a contemporary corpus: personas without a preloaded stance, demographics as their own layer, a two-sided briefing.

  • Best result: within about 9 points of the real outcome — documented and published as a measurement.
  • The cleanest finding: without counter-arguments even directly affected groups mostly agree; with deliberation they flip — self-interest only acts through weighing arguments.
  • The influence of the base model is measurable — we document it openly so results stay interpretable.

Scale experiments

Wisdom of the crowd

The classic estimation experiment, agent edition: 1,000 to 10,000 agents estimate, discuss, revise — calibrated against known ground truth. Measured in the experiment: ~3 cents per run.

Town vote

5,000 citizen personas with demographic profiles vote on a four-day week, a congestion charge and school uniforms — polarisation and segments included. ~6 cents per run.

Prediction market

10,000 traders at an order book: price discovery emerges, the LLM only reads sentiment — the market mechanics are deterministic code. ~45 cents per run.

Personal assistants — on your server

Six ready-made scenarios for daily life, all with the same pattern: read a lot, write little — and an approval before every irreversible action.

Inbox zero

Classify, archive, draft replies — nothing is sent without your approval.

Morning briefing

Calendar, weather, feeds — one push message at 6:30. The audit log shows exactly one outgoing action per run.

Tasks via messenger

"Reminder Tuesday 2 pm" — scheduled on your own cron, listable and cancellable.

Research agent

Reads, summarises, remembers — later recall against your own vector store.

Family helper

Drafts greetings and reminders — never sends without approval; sensitive data is declared as such.

Voice note to action

Local Whisper, local Llama, local database — fully offline, nothing leaves the box.

Get in touch