CommandAGI · commandagi.com/for/zero-token-agents

CommandAGI
Abrir a bancada de trabalho
Menu
Abrir a bancada de trabalhoEntrar

Zero-token agents

Plans, facts and skills, no tokens.

When the steps are known, an agent needs no model. A skill file replays them and checks each one, and a model is asked only from the first step that fails.

Open the workbenchSee decision models

Replay, check, then ask

A skill is a file of checked steps.

The skills harness turns a run that succeeded into a skill file: a phrase, its parameters, and each computer_act step with what the screen must show after it.

How a zero-token agent worksA run that succeeded is distilled into a skill file, a phrase with parameters and computer_act steps, each with an expectation. A new goal that matches the phrase replays the steps with no model call, checking each expectation on the screen. Only at the first step whose expectation fails does a model plan the rest, and the skill file is patched with what worked.A run that succeededA skill filesteps with expectationsReplayno model callAsk a modelonly from the first failedstepHow a zero-token agent worksA run that succeeded is distilled into a skill file, a phrase with parameters and computer_act steps, each with an expectation. A new goal that matches the phrase replays the steps with no model call, checking each expectation on the screen. Only at the first step whose expectation fails does a model plan the rest, and the skill file is patched with what worked.A run that succeededA skill filesteps with expectationsReplayno model callAsk a modelonly from the first failed step

Case study 1 · a simulated desktop

The same task: 0 calls against 17.

The sim_save task on a simulated desktop. The scripted agent sends the actions a correct agent sends; the baseline sends each screenshot to a vision model and takes its one action.

passed by the scripted agent, 0 model calls, $0, 2.4 s in all
3 of 3
passed by the vision-model baseline, 17 calls, $0.0013, 124 s
1 of 1

Simulated

Wall time for the sim_save task
Os números
Wall time for the sim_save task
AgentWall time
Scripted2.4 s
Vision-model baseline124 s
ProvenanceFile test-results/computer-eval/scripted-serverless-2026-10-28T06-27-57-260Z.json and baseline-serverless-2026-10-28T06-30-05-354Z.json (this computer)Output case-studies/computer-evals/stats.jsonTool scripts/workbench/computer-eval.mjs, read by scripts/sites/case-studies.mjs computer-evalsRan on This computer (dell-gb10-1), a simulated desktop (ComputerWorld)When

Case study 2 · skills on ten tasks

Skills pass as often, in a quarter of the time.

Ten desktop tasks, two runs each. The skills agent replays what it learned and asks a model only where a step fails.

Not zero yet. On these ten tasks the skills agent still called a model where a step failed. The runs with no model call at all are the scripted agent's, above.

passed by the skills agent, with 114 model calls and 443 s
11 of 20
passed by the vision-model baseline, with 282 calls and 1,729 s
10 of 20
for the skills agent (deepseek-flash) and $0.019 for the baseline (qwen3.7-flash): fewer calls, but not cheaper here
$0.057
ProvenanceFile test-results/computer-eval/skills-2026-10-25T01-49-30-309Z.json and baseline-2026-10-25T02-39-16-279Z.json (this computer)Output case-studies/computer-evals/stats.jsonTool scripts/workbench/computer-eval.mjs --harness skillsRan on This computer (dell-gb10-1), a private Xvfb desktopWhen

Teach it once, replay it free.

Run a task once with a model, and the skills harness keeps the steps.

Open the workbenchSee decision models