Agents and automations

An agent is a saved worker with instructions, knowledge, tools, and spending limits. An automation gives that agent a recurring task. You can also run it once or call it from your app. Start with Agents → Create agent, write your instructions, and add your own knowledge and tools.

The five actions

  • Build: name the agent, describe its job, and add optional knowledge and HTTP tools.
  • Try it: give it one task and inspect the result before automating it.
  • Automate: choose a task, frequency, time, and time zone; pause or edit it later.
  • Website & API: create an expiring scoped key and copy the server-side example.
  • History: find results, errors, and costs from manual, scheduled, and app requests.

You can configure agents before billing setup. Running them requires billing setup, credits, and an available spending budget. Organization members can view; owners and admins can change settings and start runs.

Uploading knowledge

Use UTF-8 TXT, Markdown, CSV, or JSON files up to 100 KB each. An agent accepts up to 10 sources and 200 KB of configuration in total. Export PDF and word-processing files to text first. Web pages must be public HTTPS text or JSON and are refreshed on each run. Uploaded files change when you replace them. Each run keeps a snapshot of the knowledge it used; deleting a source affects future runs.

Schedules and results

Choose hourly, daily, weekdays, or weekly. Times use the selected time zone. Missing daylight-saving times are skipped, repeated times run once, and after downtime only one missed occurrence is caught up. Results appear in History and your dashboard notifications. Reports are not automatically emailed or sent to connected apps.

Every run starts with fresh conversation context. For comparisons or a continuing conversation, supply the earlier data in the input. Tools can read configured public endpoints; POST tools can send JSON and should be configured only for actions you intend to automate.

If a run stops

Check its error, credits, and spending limit. Runs have bounded execution time and tool steps. A cancelled or interrupted run may have already used model tokens or contacted a tool; review those effects before starting a new run. Unknown external outcomes are never silently repeated. Recorded model usage is charged once, with settlement retried independently if needed.

Custom code and connected reviews

Agents → Custom code retains repository imports, custom runtimes, repository schedules, and connected pull-request reviews. These have their own deployment and review history. Existing repository starters keep their original code and licenses; new agents can be built from scratch without a repository.

See the Agents quickstart, workspace and app API, and connected code reviews.