Getting Started

This guide takes you from zero to a completed task: create a project, give it work, watch it execute, and respond when it needs you.

1. Create a project

Open Agent Ops in Studio and create a new project. Three things matter at creation time:

• Name and description — the description doubles as context for the agent, so describe the project's purpose the way you would brief a new team member.
• Prompt — the project's system prompt. This is standing guidance for every task: who the agent works for, what conventions to follow, what to avoid.
• Model — the AI model that powers every run in this project.

The project provisions a persistent sandbox on first use — you don't manage it yourself. Files the agent creates stay in the project across tasks and runs.

2. Configure what the agent can reach (optional)

Under the project's Configure tab you can set up, at any time:

• Actions — integrations from the gallery (GitHub, Slack, SQL databases, CRMs, email…). See Actions & Integrations.
• Skills — reusable instruction packs for procedures and formats. See Skills.
• Templates — reusable task definitions.
• Automations — webhook and integration triggers that create tasks automatically.

Under Settings, add environment variables (for example a GITHUB_TOKEN for private repositories) and review the runtime safety switches. See Projects & Sandbox.

None of this is required for a first run — a bare project can already execute code, work with files, and search the web.

3. Create and run a task

Create a task with a title and a description of the outcome you want. Be concrete about the deliverable:

Analyze the attached sales-export.csv. Produce an Excel workbook with
one sheet per region, a summary sheet with month-over-month totals,
and a chart of the top 10 products. Flag any rows with missing prices.

Run it. The task screen streams everything live:

• Plan — the agent's todo list, updated as it works
• Activity — every step, tool call, and log line
• Output files — deliverables as they land in the project

4. Respond when the agent needs you

Two things pause a run:

• A question — the agent needs information or a decision, and the run parks as waiting for input.
• An approval — the agent wants to do something gated by policy (push to git, create a subtask, call a risky action), and the run parks as waiting for approval.

In both cases the task owner gets an email with a deep link to the exact task — you don't need to keep the screen open. Answer or approve, and the run resumes exactly where it stopped. Completed and failed runs send an email too.

5. Collect the results

Deliverables appear under the project's Files tab and on the task's output panel. Because the sandbox is persistent, follow-up tasks can build on previous outputs:

Take last week's regional workbook and extend it with this week's export.
Keep the same formatting.

Where to go next

• Put a task on a schedule → Tasks, Runs & Scheduling
• Trigger tasks from external systems → Automations & Triggers
• Teach the agent your team's procedures → Skills
• Browse ready-to-build scenarios → Use Cases