Agents

An agent is a helper you build in plain words. You give it a name, say what it is for, write its instructions, choose the tools it may use, and decide which of those tools must wait for your approval. Give it a goal and it works through your tasks, projects, calendar and notes one step at a time on your workspace's AI provider, showing every step with the reasoning the model gave, what it asked for, what came back, and what the step cost, and ending with an answer. It can also run on a schedule, and you can give it a monthly spending limit. An agent acts with your permissions and never more: it sees and changes only what you could.

• Start from blank or from a template: daily planner, overdue sweeper, weekly recap, notes to tasks, ask your workspace and project check-in
• The tools an agent may use: search the workspace, read your tasks, read one task, see projects, read today's calendar, create tasks, change tasks and write a note
• Tools that change data wait for your approval by default, and the run shows exactly what the agent wants to change
• Every run is recorded: the goal, each step with its reasoning, tokens, cost and duration, its result, the final answer or why it stopped, and the total cost of the run
• An optional monthly spending limit for each agent, in US dollars: once it is reached, new runs are refused until next month
• Schedules every day, every weekday or once a week at a time in your own time zone
• Pause an agent to stop its schedule without losing it, or archive it to remove it from your list

• An agent runs only when your workspace has an AI provider. Without one, the page says so and an owner or administrator can add a key in Settings, AI providers.
• An agent acts with the permissions of the person it runs for. It cannot see or change anything that person cannot.
• A run stops after twelve steps. If the goal needs more, give the agent a narrower goal or split it into two runs.
• A step waiting for approval holds the run until somebody approves or rejects it. Rejecting ends the run.
• A scheduled run uses the goal you wrote for the schedule, and runs as the person who built the agent.
• Runs use your AI provider and count toward its usage and cost. A run counts toward the agent's monthly spending limit when it finishes. At the limit, new runs, scheduled ones included, are refused until the next calendar month (UTC) or until the limit is raised.

1. 1Open Agents and choose New agent, or Use template on a template.
2. 2Name the agent, say what it is for, and write its instructions in plain words.
3. 3Choose its tools. For a tool that changes data, keep Ask me before it runs on unless you trust the agent to act alone.
4. 4Optionally enter a Monthly spending limit (USD), and choose when it runs on its own, the time, and the goal for those runs. Then choose Create agent.
5. 5On the agent page, write a goal and choose Run. Each step appears as the agent takes it.
6. 6When the agent waits for approval, read what it wants to change, then choose Approve to let it continue or Reject to stop.

What can an agent do?

Only what its tools allow: search the workspace, read your tasks or one task, see projects, read today's calendar, create tasks, change tasks and add to your note for a day. It cannot send email, delete anything or reach other apps, and it acts with your permissions.

Why does Run stay unavailable?

An agent needs an AI provider for your workspace. The page shows what is missing. An owner or administrator adds a key for Anthropic, OpenAI, Gemini or Azure OpenAI in Settings, AI providers. A paused agent also cannot run until you resume it.

How do I stop an agent changing my tasks without asking?

Keep Ask me before it runs on for Create tasks, Change tasks and Write a note. It is on by default. The agent then pauses at each change and shows you exactly what it wants to do.

Why was my run refused with a message about a limit?

The agent has spent its Monthly spending limit for this calendar month (UTC). Edit the agent to raise the limit or leave the field empty, or wait for the next month. Runs that are still working are counted when they finish.

Who can open Approvals and governance?

Workspace owners and administrators. Other people can build and run their own agents, and decide the steps their own runs wait on, from the run itself. An owner or administrator can decide a waiting step from the Approvals queue as well: approving it there runs the step and resumes the run, exactly as approving it on the run page does. The Agent templates page lists the available templates for reading only; to build from a template, choose Use template on the Agents page.

What happens to a run if Atlas restarts during it?

A run that has made no progress for fifteen minutes is marked as failed, with a note that it was interrupted before it finished, and you can start it again. Steps it already finished stay recorded.