Chatting with an agent: tasks

Gumloop · Essentials · 7 min

Chat is how you test, use and correct every agent, so it's the first skill to get comfortable with.

You use an agent by chatting with it, much like a messaging app. Each piece of work you start is a task: it has its own conversation, its own files and its own cost.

In the message box, the + menu lets you:

  • Add photos & files, such as a spreadsheet to analyse.
  • Use skill: point the agent at one of its playbooks.
  • Mention integration: nudge it to use a specific app.
  • Start an Incognito task, whose messages are not kept in task history.

You can also speak instead of typing; your voice is turned into text.

While the agent works you see the steps it takes (searching, calling an app, writing a file). You don't have to wait: send a follow-up message to add details, or use Send Now to redirect it immediately if it is heading the wrong way.

A context meter shows how much of the model's working memory the conversation is using. Long conversations fill it up, cost more, and can make the agent lose track of early details. For a new topic, start a new task instead of continuing an old one.

The agent may stop to ask you something, for example with clickable choice cards, or ask permission before an action (approvals come later).

The Tasks page lists all your work across agents. Filter by date, agent, source (chat, trigger, Slack...) or status, and rename, share or delete tasks. Each task shows its running credit total.

Remember

  • Each piece of work is a task with its own conversation, files and cost
  • Steer a running agent with follow-ups or Send Now
  • Start a new task for a new topic to keep context small

Check yourself

  1. Your long task conversation is getting slow and the context meter is nearly full. You now want help with an unrelated topic. Best move?

Sign in to answer, see the explanations, and get this lesson back as a quick review later.

Official docs: docs.gumloop.com/core-concepts/agents