CommandAGI · commandagi.com/for/decision-models

CommandAGI
Abrir el banco de trabajo
Menú
Abrir el banco de trabajoIniciar sesión

Decision models

One choice, one pass.

A decision model holds no conversation and calls no tools. It reads a state and a typed question and answers in a fraction of a second. CommandAGI makes each one an agent.

Open the workbenchSee zero-token agents

Driving a computer

A table of controls, one pick each step.

The jev-table harness reads the screen into a numbered table of controls and facts. One Jev request picks the action, the item and the key. DeepSeek writes free text and takes over when Jev is unsure or stalls.

How a decision model drives a computerEach step, the harness reads the computer's screen and accessibility tree into a numbered table of its controls and computed facts. One request to the Jev decision model picks the action kind, the item, the key and the place. The driver acts on the computer. When Jev's confidence is low or the task stalls, a DeepSeek model writes the text or recovers.The screena private Xvfb desktopA numbered tablecontrols and factsThe decision modeltypesafe/jev-1.13The actionclick · type · key · dragHow a decision model drives a computerEach step, the harness reads the computer's screen and accessibility tree into a numbered table of its controls and computed facts. One request to the Jev decision model picks the action kind, the item, the key and the place. The driver acts on the computer. When Jev's confidence is low or the task stalls, a DeepSeek model writes the text or recovers.The screena private Xvfb desktopA numbered tablecontrols and factsThe decision modeltypesafe/jev-1.13The actionclick · type · key · drag

Case study 1 · 24 computer tasks

40 of 48 attempts, for $0.14.

The computer-use eval's 24 tasks, two runs each: save, type, drag, dialogs, forms, a terminal, settings and more, on a private desktop at 1920 × 1080.

Not the same test as a vision model's. The vision-model baseline (baseline-2026-10-25T02-39-16-279Z.json) ran 10 of these tasks: 10 of 20 passed, with 282 calls, $0.019 and 1,729 s. The task sets differ, so this is not a ranking.

of 48 attempts passed (83%)
40
Jev decisions and 128 DeepSeek calls
255
for all 48 attempts
$0.142
of wall time, 365 s of it in the models
652 s
ProcedenciaArchivo test-results/computer-eval/jev-table-2026-10-25T21-49-18-102Z.json (this computer)Resultado case-studies/computer-evals/stats.jsonHerramienta scripts/workbench/computer-eval.mjs --harness jev-table, read by scripts/sites/case-studies.mjs computer-evalsSe ejecutó en This computer (dell-gb10-1), a private Xvfb desktopCuándo

Case study 2 · one decision, every model

A decision in a tenth of a second.

Every decision model OpenRouter's decisions door and each publisher's own door served on 2026-10-25, asked the same decision three times from this computer. Latency is the whole request; cost is what the route billed. The whole check cost $0.0012.

Each model had one input here, so this says nothing about accuracy.

median, TypeSafe Jev 1.14, at $0.000018 a call
0.10 s
median, Microsoft-Decision-1
0.33 s
median, Upstage Solar Decide: an order of magnitude slower on this input
10.6 s
Median time for one decision, by model
Los números
Median time for one decision, by model
ModelMedian
Jev 1.140.1 s
Elephant-10.16 s
GPT-6 Luna Decisions0.24 s
Liquid D10.26 s
Clef0.32 s
Microsoft-Decision-10.33 s
Kev 4B0.63 s
Span-028.7 s
Solar Decide10.6 s
ProcedenciaArchivo docs/system-one.md § measured for a decisionHerramienta three calls of each model, decide with one state and three questionsSe ejecutó en This computer (dell-gb10-1), each model's routeCuándo

Give an agent a fast choice.

Pick a decision model as an agent, or as the chooser of another agent's next step.

Open the workbenchSee zero-token agents