Skip to main content
Agent Studio is the composer for starting managed workers. Open it from Agent Swarms (Observe → Agent Swarms → + New Agent). You land in the agent gallery first—pick a template or start from scratch—then configure the run in the Studio slide-over. The gallery groups starting points by intent:
SectionUse for
Start from ScratchOpen Studio with an empty task—you choose preset, project, skills, and prompt
Run a SessionOne-off investigations or repo fixes with a seeded prompt
AutomationsRecurring health checks or metric-triggered triage
Built-in session templates include Investigate failing traces and Fix a bug in your repo. Automation templates include Triage a monitor alert and Recurring health check. Each template pre-fills the task prompt and highlights the controls you should set next (project, preset, skills, or automation name).
Agent Studio agent gallery grouped by intent with Start from Scratch, Run a Session, and Automations templates

​
Session vs automation

ModeUse for
SessionOne-off work; follow the transcript and send follow-ups
AutomationRepeating task on a cron schedule or metric threshold
Sessions can run with a saved agent preset or with None to configure harness, sandbox, and skills for that run only. Automations require a preset—they inherit harness, sandbox, repo, and skills from it. Switch to automation mode with Setup as automation in the composer (available when an Arize sandbox preset is selected; Claude Managed Agent sessions are session-only). Signal uses the same automation model on each LLM project.

​
Configure a run

1

Pick a preset (optional)

Choose a saved agent preset, or None to configure manually. Built-in Signal presets are not listed here—they are managed separately.
2

Select project

Traces and evals are read from this project for every run in the session or automation.
3

Attach skills and repo (when needed)

With None selected, add skills inline. With a preset, skills and repo come from the preset (edit the preset to change them).
4

Enter task prompt

Describe the investigation or output you need—for example, root-cause a tool failure or draft eval labels for faithfulness failures.
5

Run session or create automation

Click Run session for a one-off job, or Setup as automation → configure trigger and name → Create Automation.
On first run you may need to acknowledge that Claude Code can process project traces, your instructions, and connected repo or tool data.

​
After you start

When a session starts, Studio closes and Agent Swarms opens the Agent List tab with the new run’s detail slide-over. The worker provisions a sandbox, clones the repo (if configured), loads skills, and pulls project context. From session detail you can:
  • Read the live transcript
  • Send follow-ups to steer the run
  • View traces for the sandbox run
  • Reprovision a completed or stopped session to continue work
Manage all runs from Agent Swarms.

​
Outputs

OutputWhen
Investigation markdownAlways possible
Eval labels, dataset rowsTask-dependent
Branch / PRPreset or run includes repo write access
Review and merge PRs through your normal GitHub workflow. These pages sit under Managed Agents in the docs sidebar: