Choose the right capability
Ask what the task is missing before choosing a capability. Do not combine advanced-looking features by default.
| What you need | Start with | Why |
|---|---|---|
| Read, search, or modify the current project | Ordinary Tool | Existing file, command, and Git capabilities are sufficient |
| Reuse a fixed workflow, template, or domain guidance | Skill | Package instructions and resources for task-specific activation |
| Call tools supplied by a specialized local system | MCP Server | Add structured external Tools for the Agent |
| Navigate or act on a real webpage | Built-in browser automation | Use a managed browser with dedicated risk approvals |
| Run several independent investigations in parallel | Multi-Agent | Subagents divide the work and the root Agent synthesizes it |
| Repeat the same task on a schedule | Scheduled Automation | A persistent Task starts an ordinary root Run on schedule |
| Get one short answer | Ordinary conversation | Least configuration and lowest coordination cost |
Common combinations
Skill + Tool
A Skill defines how to work, while an ordinary Tool reads files or creates an artifact. For example, a document Skill guides the structure and a file Tool writes the output.
Skill + MCP
A Skill defines the team's workflow while MCP provides an interface to a specialized system. Each remains subject to its own permissions and approvals.
Multi-Agent + ordinary Tools
The root Agent divides an investigation into frontend, backend, and testing workstreams, and each subagent uses the same foundational Tools. This works well for independent evidence collection, but not for several workers modifying the same file concurrently.
Scheduled Automation + Multi-Agent
A scheduled Run starts a root Agent, which can then create subagents as needed. Scheduled Automation cannot bind directly to a subagent conversation and does not provide additional concurrency guarantees.
Names that should not be confused
- Scheduled Automation starts an Agent Run on a schedule.
- Browser automation lets an Agent operate a managed webpage.
These are separate subsystems. One answers “when should this run?” and the other answers “how should the webpage be operated?”
The least-capability principle
The smallest capability that can complete a task is usually the easiest option to verify:
- If a read-only file Tool is sufficient, do not run an arbitrary command.
- If an ordinary prompt is sufficient, do not create a permanent Skill.
- If one Agent can complete the work, do not divide it among subagents.
- If recurring execution is not required, do not create a Scheduled Automation.
- If you do not understand the behavior of a third-party MCP Server, do not enable automatic calls.
For the current support boundaries of each capability, see Capability limits.