Workflows · AgentInputLab Studio

Configuration generation is available. To enable inspection and execution, restart with agentinputlab-studio --workflow-root ./studio-workflows.

Loading available profiles…

Choose a starting point

Presets create editable drafts. They do not supply a model service or private artifacts.

Preset Exact method profile Benchmark Number of cases Query for input-only runs

Connect your model

Leave all model fields empty for a prompt-only run. Model weights are not downloaded by this form.

Served model ID API base URL Model / deployment revision API key environment variableSet its value in the server shell; never paste a key here.

Set per-case budgets

Concurrency changes scheduling. Each case keeps its own token and call limits.

Total output tokens Model calls Retrieval calls Tool calls Seed Case workers

Method parameters, retrieval and runtime settings

Use the method's native parameter names for temperature, per-call output caps and max_length. Unknown parameters are reported by preflight. File paths are on the server.

Method parameters (JSON) Retriever (JSON) Artifacts (JSON) Native runtime (JSON) Preparation (JSON) Simulator / judge backend for arenas that require it (JSON)

Existing run.json can be pasted into the run object below. Extra target/scorer settings, case IDs, cache paths and local-model settings are editable in the full workflow.

Import workflow.json or run.json

Review and run

This JSON is the execution configuration. Editing it does not start a model or download.

Complete workflow.json

Generate or import a configuration to begin.

Selected method's parameter reference

Generate a configuration to show the native parameter definitions.

Check model endpoint availability (no generation)

Allow configured dependency / dataset downloads during preparation

Running a model-backed workflow sends requests to your configured model service. Native scores, failures and consumed budgets are recorded.

Saved runs

Refresh or reopen this page to inspect persisted runs. Resume uses the saved configuration and does not repeat completed attempts.

Full execution and preparation report

Outputs and native scores

Bounded preview of the first 20 terminal records; the complete journal stays in the run directory.

Native records and provenance