How the farm works
This is not a pitch deck. It is a description of the system that produced this page — the same one that runs our client work. Agents do the work; a human directs and reviews; the pipeline ships.
One human, holding the thread
A single operator sets direction, reviews what matters, and signs off on what ships. There is no middle layer to brief, chase, or wait on. The leverage of an agency without the agency — because the queue is worked by agents, not headcount.
Agents on isolated worktrees
Work runs in parallel on isolated git worktrees — each agent on its own branch, its own checkout, its own task. They don't trip over each other, and nothing touches the main line until it's reviewed. This page was built that way, in a worktree, off main.
Gather, draft, validate
Content that comes off the farm — field notes, the operations feed — runs through a strict three-step sandwich. A deterministic gather step collects only the facts. Exactly one model call phrases them. A deterministic validator is the gate: it hard-fails on anything that could identify a client, then writes what publishes. The model phrases; it never decides what is safe to say.
Signed, reviewed, then it lands itself
Every commit is cryptographically signed. CI runs the checks. Pull requests are reviewed — and once a reviewed PR is green, it merges itself. The human decision is the review, not the button. Receipts, not adjectives: the merges below are real, public, and linkable.
MCP-first contact
The cleanest way to reach the farm is the same protocol agents speak. The site exposes an MCP endpoint an assistant can connect to directly; a form and email are there for everyone else. Start a conversation the way the work actually happens.
https://rarebit.one/mcppublic PRs merged across the org in the last 30 days
total downloads of the open-source gems the farm runs on
this page rebuilt at, in farm-local time
- rarebit-static-v3#233Field notes: rate-limiting-and-ci-cleanup-across-repositories14 Jul
- standard_id-google#66chore(deps): bump the actions group with 2 updates14 Jul
- standard_health#39chore(deps): Bump the actions group with 2 updates14 Jul
AI woven into every workflow
Agent-first delivery
Every engagement starts by mapping which workflows an agent can own end-to-end — and which need a human in the loop.
Human-in-the-loop by design
Your tools, not new ones
We plug agents into the tools teams already use. The same automations that run Rarebit run our client work — no new platform to adopt.
The same farm, pointed at your work
Everything above runs Rarebit. None of it is bespoke for this site. When we take on client work, we point the same agents, the same worktrees, the same pipeline, and the same review gate at your queue — inside the tools your team already uses.