Documentation › The console

The console

The console is the normal way to work. It is at your instance’s address — for example https://starcraft.staging.thalamuscua.thalamusaicloud.com, the instance starcraft in its staging environment. Its pages follow a goal’s life: write it, run it, step in when needed, and review what happened. Every page has a ? beside its title that explains what the page is for and how to use it.

Home

Where to start: what is running, what needs you, recent runs, and short explanations of how ThalamusCUA works, how to use the console and the safety built in.

New goal

Pick the application and the environment (a note shows the address and login it will use), choose the mode, and write what you want done in plain words. Only a disposable environment allows Supervised or Autonomous; everywhere else the run is Read only.

Press Start run. If the goal is ambiguous, the agent asks first (“Before I act, please clarify”) rather than guessing; answer and press Answer and continue. As soon as planning finishes the run starts and its run page opens.

How-to: give the agent a goal →

Plan editor

The tasks the planner made for a goal, each with its risk level, skill and checks. Reorder them with ↑ and ↓, rewrite what a task should do, or Remove it; then Save draft or Approve & save. Each save is a new version of the plan.

New goal starts the run as soon as planning finishes, so an edited plan is for runs started afterwards; a run already going keeps the plan it started with. The goal id is in the Goal column of Runs.

How-to: review and adjust a plan →

Runs & run detail

Runs lists every execution: its id, goal, description, benchmark, date and state. Open one for its run page, which updates as the run works:

• a stage banner saying what the run is doing now (planning, working, waiting for you, failed) and, for a failed run, the likely cause;
• Answer (the result and the screen it was read from), Timestamp (the activity log) and Steps (each action with its screenshot and reason);
• Approval cards for actions on this run that wait for a decision, and a Send a hint… box while it is running;
• a playback of a finished run’s recording, Trace explorer & report export, and Cancel run (it asks you to confirm).

How-to: watch a run →

Approvals

High-risk actions the agent has paused on, waiting for a person. Approve to let it continue, or reject to block that agent’s work — other agents in the same run carry on.

How-to: step in while it works →

Trace explorer

Every event of a run, filterable: step through the agent’s observations and actions with their screenshots, and download the report (readable Markdown or structured JSON).

Application briefs

What the agent knows about an application. Generate draft has a large vision model draft the brief from the application’s screens; you edit it and press Approve & commit. Only an approved brief becomes trusted context. The same page lists selector-healing proposals — suggested fixes for controls the agent can no longer find — to accept or reject.

How-to: review what the agent knows →

Environments

Each instance the agent works in: its address, the login it uses, whether it is disposable, and snapshot and reset for disposable ones. See Environments & safety.

Benchmarks & Evaluations

Test cases for each application, and runs of them against a real instance or the mock site. See Benchmarks & evaluation.

Configurator

Which model does which job, and where the model servers are. Pick a model that has been measured and a note appears under its role with the results, before you save. See Models.

Documents

The model evaluations — each one’s report, summary and raw results — and the benchmark scripts and task set that produced them. Evaluation and Scripts tabs list them ten at a time; click one to open it in a window with Source, Copy path and Download. The Configurator’s note links straight to its evaluation here.

How-to: compare models before choosing one →