Skip to content
Help Center Back to AgenticObjects

Agent Studio

The agent inventory and the Agent Studio — create an agent, give it instructions, choose the data it may reach, set its budgets and schedule, then publish it.

Console 5 min read Updated

The Agents page lists every agent in your workspace by lifecycle state. Opening an agent takes you into the Agent Studio, where everything that shapes how it answers — its instructions, its model, its data, its budgets and its schedule — is edited one section at a time.

What this screen is for

An agent is a governed analyst: a persona, the data models and metrics it may reach, the budget it may spend and the rhythm it runs on.

Every agent follows the same lifecycle. It is created as a draft, which never runs and answers nobody. When it is ready you publish it, and it becomes a stopped agent. You then activate it deliberately. That separation means an agent reaches your colleagues only when someone has chosen to let it.

What you see

  1. Agents — the page title, Domain-calibrated agents over your governed data models beneath it, and + New Agent. The top bar offers Search agents; the workspace and project are chosen at the top of the left menu.
  2. Active, Stopped, Drafts and Archive — one tab per lifecycle state, each with its count.
  3. Table / Cards — two views of the same inventory. The table shows Name, Status, Owner, Last run, Schedule, Distribution and Project, grouped by project; the cards carry the same signals. Each row offers its next step: Activate, Deactivate, Unarchive or, for a draft, Continue.
  4. Create an agent — a page of its own with two steps: Pick a start (a template card or Start blank) and Name it and choose its data.
  5. The Agent Studio header — the agent's name, its type, its status, the owner, the bound Model links and a Project selector.
  6. The section rail — General, Instructions & Personality, Data & Scope, Output, Limits, Schedules & Distribution and Run History. Exit studio returns you to the inventory.
  7. The panel strip — each section's panels as buttons: Overview and LLM; Persona, Orchestration, Tools and Prompt Preview; Data; Output; Guardrails; Schedule, Run Now and Distribution.
  8. Save this block — each panel saves on its own. Panels that change how the agent answers say so: This changes how the agent answers, and it applies from the next run.

How to …

Create and publish an agent

  1. Press + New Agent.
  2. Under Pick a start, choose a template — it prefills persona, brief format and guardrail suggestions — or Start blank.
  3. Under Name it and choose its data, enter a Name, choose the Workspace, Project and Data model, and set the Metric Scope: Every metric in this model or Only the metrics I choose.
  4. Press Create draft. The Agent Studio opens with a checklist of what is still missing, each item with a Go there link.
  5. When the checklist is clear, press Publish, then Activate when you want the agent to start working.

Decide which metrics an agent may reach

  1. Open Data & Scope.
  2. In Metric Scope, search the model's metrics, tick the ones you want, or use All and None.
  3. Read the line above the list — for example Scope: 12 metrics selected.
  4. Press Save scope. To bind a further data model, use Add a data model.

Set the budgets a run may spend

  1. Open Limits.
  2. Edit the Scheduled budget and the Interactive budget separately — steps, time, cost and the other limits each profile uses.
  3. Leave a field empty to inherit the workspace value; a value you set is marked agent-specific.
  4. Press Save this block.

Check what the agent will actually be told

  1. Open Instructions & Personality, then Prompt Preview.
  2. Read the compiled prompt, section by section, with its token count.
  3. The preview is read-only and exact: Read-only — this IS the compiled prompt the run executes.

How to read it

  • The Output section shows each type on its own. Summary, Findings and Recommendations each have a row: whether it is produced, how many at most, and whether it is published on its own. The rows show the agent's saved setting; when it matches a ready-made shape, its name is written below. Findings not published on their own are read only inside the summary.
  • The channel changes the budget, never the authority. Chat and scheduled runs can have different budgets, but the data and tools an agent may use are the same set of rules.
  • Changes apply from the next run. A run already in progress keeps the settings it started with; use Run Now to see a change straight away.
  • The model panel says exactly what serves each channel. A line such as Chat: … · Scheduled runs and Run Now: … names the model and where the choice comes from. Use a different model for chat lets chat run on a faster model than the night job.
  • Only the controls a model accepts are shown. Answer variability and Thinking appear when the chosen model supports them.
  • Tools are shown per mode. Conversation, Scheduled and Compare modes show which tools each channel receives; required tools are always on, and any tool that is withheld says why.
  • Lessons carry their state. Each lesson is badged awaiting approval, approved, ⚡ auto or reverted, and can be approved, reverted or dismissed. What will this lesson sit beside? shows the instructions it will join.

Good to know

  • Nothing is ever deleted. Archiving an agent stops its schedule and delivery; the agent stays readable under Archive and can be restored with Unarchive.
  • Budgets and data narrowing are tighten-only: you can restrict below the workspace or model setting, never loosen it, and the engine checks every save.
  • The serving model changes only when a person changes it. A model change is an audited decision that applies from the next run.
  • The distribution list is managed by workspace admins. Followers choose an agent for themselves, so their list is shown here but never edited on their behalf.
  • Data models are made in Model Studio (built there, or imported there from a model file). The Agent Studio binds them to agents and narrows what each agent may reach.