the rig that runs the loop
The lathe of the intellectual revolution.
GPUs are the new steam engines. Neural networks are the new power. looprig is the machinery that turns that power into dependable, governed, useful work: an agent harness of stdlib-first Go modules, for agents you own and can change.
go get github.com/looprig/harness | the industrial revolution | the intellectual revolution | |
|---|---|---|
| the engine | the steam engine | GPUs and TPUs |
| the power | mechanical power from heat | knowledge and reasoning at scale |
| the machinery | the lathe, mother of machine tools | looprig, the agent harness |
| what it makes | precise, repeatable parts for other machines | durable, confined, gated, transparent work |
The steam engine supplied the power. The lathe made the power useful.
- 3.1 loop
Defines one agent: its model, instructions, tools, and permissions.
- 3.2 rig
Assembles one or more Loops with storage and lifecycle policy.
- 3.3 session
One running instance of a Rig. Submit work, observe events, answer gates, resume after process death.
test record t-014 / durability certification
$ coderig > migrate billing tests to testify ▸ session 019823a1 · turn 1 · harness/loop ▸ tool: bash · go test ./billing/... ◼ gate: bash wants write access · waiting for a human ✓ approved · gate recorded in journal ✗ process killed (kill -9) $ coderig --resume 019823a1-7b3c-4d5e-8f90-123456789abc ✓ journal replayed · 14 events · nothing lost ▸ turn 1 resumes exactly where it stopped
- 01
durable
Work survives crashes, restarts, and long pauses.
Every turn is journaled as it happens. Sessions suspend, restore, and resume across process death, proven by an external cross-repo suite against a real store.
- 02
confined
Intelligence acts only within explicit boundaries enforced by the OS.
Not prompt-level guardrails: Seatbelt on macOS; namespaces, Landlock, seccomp, nftables, and cgroups on Linux. What the platform cannot enforce is reported honestly.
- 03
gated
Human judgment stays present wherever consequence demands it.
Approval gates are a durable primitive, not a UI feature. A gate can be answered minutes or days later, even after a restart, and the loop resumes exactly where it paused.
- 04
transparent
Every consequential action is observable, explainable, and auditable.
An append-only journal is the source of truth for everything the agent did: what it saw, what it ran, who approved it, and what it cost.
- 05
independent
The rig outlasts models, vendors, machines, and environments.
A provider-neutral inference contract, storage behind small interfaces, and OS confinement with no framework lock-in. The engines change. The lathe remains.
These are not accessories added after an agent becomes capable. They are the precision machinery that makes capability useful.
surfaces
svelte / react / vue / angular / solid / plain ts, one client core
the heart
rig partners
fig. 2 one harness, replaceable parts
| part | your choice |
|---|---|
| models and providers | any provider behind one neutral contract; local, cloud, or TEE |
| instructions and knowledge | yours, versioned with your code |
| tools and permissions | select standard tools one at a time, or write your own |
| storage | single-host disk, NATS JetStream, or any object store via rclone |
| interface | terminal TUI, web, or fully headless |
| where it runs | laptop, server, or fleet; one static binary |
Production-grade does not mean enterprise-only. An agent you rely on should not forget its work after a crash, touch things outside its boundaries, act without approval when consequences matter, or hide what it did. Those properties matter for one person with one laptop as much as for a fleet.
The engines will change.
We will continue to be the lathe.