Tasks
What is a task?
In ordinary conversation, a task is the work you ask an agent to complete. The implementation records several different kinds of activity; a request, a reasoning cycle, a tool call, and a legacy browser task are not interchangeable units.
| Term | Meaning |
|---|---|
| Work request | A message or trigger asking the persistent agent to produce an outcome |
| Tool action | A recorded invocation such as a search, file operation, or email send |
| Browser task | A separate browser automation job with its own status, result, and API lifecycle |
| Inference usage | Measured model requests and tokens, with optional cost estimates |
How tasks work
An agent may use several reasoning cycles and tool calls to complete one request. It can wait for approvals, credentials, files, or asynchronous results. A plan describes intended work; it is not a billing counter or proof that every step has executed.
Example
You ask the agent to check a weather forecast, find suitable activities, and prepare a report. The agent may search, read sources, reason about the options, and create an output. Another run may need additional searches or fewer tool calls.
Those are workflow steps, not a promise of five billable tasks or five units of usage. Review recorded activity and measured usage for the actual run.
Types of tasks
| Activity | Examples |
|---|---|
| Search and browse | Search, public-page reading, and browser automation |
| Messaging | Chat replies and configured email delivery |
| Data and files | SQL queries, charts, reports, and workspace file operations |
| Code | Python or shell execution where sandbox compute is enabled |
| Integrations | HTTP, MCP, and connected-app operations |
| Planning | Plan updates and human input requests |
Tasks and inference usage
Each application-controlled provider request has a usage receipt. Tool steps and browser summaries do not add a second inference charge. Missing provider measurements remain unknown.
Use Usage for tokens, model provenance, pricing and measurement limits. The legacy browser-task API is a separate execution surface.
Intelligence tiers
Available settings depend on the deployment and entitlement. Compare result quality and measured consumption rather than treating a tier name as a capability or cost guarantee. See Intelligence Selector.
Task execution
The runtime can execute eligible independent tool calls concurrently within its policy and concurrency limits. Other work is serial or waits for an asynchronous result. Timeline entries record activity and outcomes; their display order does not prove that every external action ran sequentially.
Failures can require different responses. Missing access may need reconnecting, an approval may need a human decision, and an uncertain external write needs reconciliation before retrying. See Troubleshooting.
Agent budgets and limits
Usage is measured; token and USD limits apply only where an administrator has enabled them. License, security, pause, loop, time and concurrency safeguards remain active. Inspect measured consumption in Usage and pause an agent through Agent Settings when needed.
Monitoring tasks
Use the chat timeline to inspect tool activity, pending requests, errors, and saved results. Open Usage in the same workspace for aggregate consumption. A plan, accepted request, or queued browser task does not by itself establish successful completion.