Every command, on the record

Written down before it is sent.

An agent's command is a line in a stream before it reaches the arm, the shell or the model. The reply is another line. A reply never says the command worked unless the device said so.

Un brazo SO-101 en simulación levanta un cubo blanco de una base azul hacia una base naranja; una copia azul translúcida del brazo muestra dónde planea estar su política medio segundo después.
Recorded in the workbench: an SO-101 arm in simulation. Each of its moves is a command line in the run before the engine applies it.

The command, then what happened.

The first held-out attempt of the policy loop of 2026-10-29. Each command names who sent it; the engine's applied event follows it, with the step it landed on. The hlc and salt of each line are left out.

runs/20261029t082227z-9y8d/streams/sim-input@policy-loopreal records · a simulated run · 2026-10-29
08:22:27.934 command agent:policy-loop: grip  close=false  robot=arm
08:22:27.935 event   sim: applied  command=1  step=0
08:22:27.973 command agent:policy-loop: ik  target=(-0.113, -0.072, 0.843)  axis=-z  tolerance=0.03
08:22:27.979 event   sim: applied  command=2  step=120  error=0.0054
08:22:28.029 command agent:policy-loop: ik  target=(-0.113, -0.072, 0.835)  tolerance=0.02
08:22:28.029 event   sim: applied  command=3  step=480  error=0.00005

Case study 1 · one policy loop

742 runs, every one a folder.

The loop of 2026-10-29 recorded each attempt as its own simulated run: the commands, what the policy saw, the outcome and the judge's critique, each a stream. Counted on 2026-11-04 (case-studies/records/stats.json).

Simulated

records in 742 runs
19,178
commands in the sim-input streams, each with its applied event
8,306
policy decisions, one per attempt, each a command
734
critiques by the judge, one per attempt
734
ProcedenciaArchivo ~/commandagi-runs/cube-pick-loop-2026-10-29/runs (this computer)Resultado case-studies/records/stats.jsonHerramienta scripts/sites/case-studies.mjs recordsSe ejecutó en This computer (dell-gb10-1)Cuándo

Case study 2 · the backlog thread

1.66 M records, sealed as they are written.

In CommandAGI's own backlog thread, every model call is an intent, then a command with the agent that sent it, then an outcome. A seal (ed25519, its root anchored on Solana devnet) closes each span of lines, so a changed line no longer matches.

Unknown stays unknown. An intent with no outcome is a call whose result never came back. The record does not fill it in.

records in 174 streams
1,662,482
of them seals
608,516
intents, 6,763 commands, 6,422 outcomes in the effects streams
6,799
effects streams of monitor wakes in CommandAGI's production Drive
12,431
The backlog thread's effects records
Los números
The backlog thread's effects records
RecordRecords
intent6799
command6763
outcome6422
ProcedenciaArchivo runs/20261023t132208z-0aea (the backlog thread, on this computer)Resultado case-studies/records/stats.jsonHerramienta scripts/sites/case-studies.mjs recordsSe ejecutó en This computer (dell-gb10-1)Cuándo

Run it, and read the record.

Every run is a folder of plain files you can diff.