Documentation › Key concepts

Key concepts

Reference· 5-min read

Six terms make the rest straightforward: application, environment, goal, plan, execution and agent run. The others are here too, with where they appear in the console.

The six core terms

ApplicationEnvironments, Brief review: A system the agent knows how to operate — for example ThalamusTrace or Thalamus SRE — described by a brief and a set of skills. Onboarded once, used by many goals.
EnvironmentEnvironments: One instance of an application. Disposable: the agent may change data, and it can be snapshotted and reset. Shared: read-only to the agent; writes are refused.
GoalNew goal: What you want done, in plain words, for one application and environment.
PlanPlan editor: The tasks the planner makes to reach the goal, each with a risk level and checks. You can reorder, rewrite or remove tasks in the Plan editor; each save is a new version.
Execution (run)Runs: One attempt at a plan, with a status: running, paused, cancelled, completed or blocked.
Agent runRun detail: One agent working inside an execution. Independent steps can run as several agents at once; a lease stops two of them editing the same thing.

During a run

Skill: A verified operation an application supports. When the agent lacks one for what it must do, it asks rather than guesses.
ApprovalApprovals: A high-risk action paused until a person approves or rejects it.
QuestionNew goal: The agent asking you to clarify a goal before it plans (“Before I act, please clarify”), rather than guessing.
HintRun detail: You telling the agent something while it runs; it is used in its next decision.
TakeoverAPI: You driving the screen yourself while the agent waits; your input is recorded too. Available through the API, not yet as a console button.
TraceTrace explorer: The recorded evidence of a run: every observation, action and screenshot, hash-linked so tampering shows.

What the agent knows

BriefBrief review: A description of an application — what it is, its screens, how it is used — that the agent plans from. Drafted by a model, then approved by a person; only an approved brief is trusted.
SelectorBrief review: How the agent finds a particular control on a screen. When a screen changes, a fix is proposed and a person accepts or rejects it.

Measuring

Benchmark, suite, caseBenchmarks: A benchmark is a set of test cases for an application, grouped into suites. Each case is a goal with checks on the result.
EvaluationEvaluations: Running a benchmark (or one suite) and scoring every case — on a real instance or the mock site.
Mock site: A stand-in for the application used in tests: fast and free, but it checks the plumbing rather than the model.
Failure classificationEvaluations: Each failure is tagged by kind; a confirmed one can be promoted into the regression suite.
Model evaluationDocuments, Configurator: A measurement of vision models on one machine — calibration, the S1 tasks and step time — kept under docs/model-evaluations/. Its results appear on the Configurator when you pick a measured model.
Calibration: Asking a model to point at 12 known targets on a 0–1000 grid over the screen; 11 of 12 is a pass. It shows how accurately the model can say where to click.

Integration

Trigger: A way for another system to submit work with a service token, optionally calling a webhook when it is done.
Replay: Re-running a recorded execution to reproduce what happened.
Tenant: Everything lives inside a tenant (the local one is called home).

Goal to execution lifecycle — Goal → plan → execution → agent runs → trace.