Introduction to Agents
Agents mode transforms how Multi handles ambitious tasks. When enabled, the supervisor agent delegates work to independent subagents-each operating with its own context window and focused scope.
How It Works
Section titled “How It Works”Instead of processing everything in a single context, Multi breaks down complex work:
- Supervisor agent receives your task and creates a plan
- Subagents are spawned to handle specific portions of the work
- Each subagent completes its assignment and reports back
- The supervisor coordinates results and continues until the task is complete
This architecture means you can run tasks that would normally exceed context limits-large refactors, multi-file changes, or long-running operations-without hitting the wall.
Built-in Agents Behind the Modes
Section titled “Built-in Agents Behind the Modes”Two built-in agents back the modes:
- Quick mode runs a single general agent that handles your task end to end.
- Autonomy mode runs a supervisor agent that delegates to subagents.
Delegation is synchronous: when the supervisor delegates to a subagent, it waits for that subagent to finish and return its result before moving on. Subagents are not background or fire-and-forget jobs. Parallelism comes from dispatching multiple subagents together - see Parallel Execution for how that works.
Beyond the built-in modes, you can author your own agents - see Custom Agents. The built-in Multi Doctor agent is also selectable from the agent picker.
Subagents working on a complex task
When to Use Agents
Section titled “When to Use Agents”Enable agents for:
- Large-scale refactoring across many files
- Complex features requiring deep context in multiple areas
- Long-running tasks that would normally fill the context window
- Tasks on smaller context models that need to punch above their weight
- Multi-step operations where each step benefits from focused attention
Keep agents disabled for:
- Quick fixes and simple edits
- Single-file changes
- Tasks where speed matters more than scale
- Straightforward questions or small code generation
The Tradeoffs
Section titled “The Tradeoffs”Agents mode is powerful, but comes with costs:
| Aspect | Agents Enabled | Agents Disabled |
|---|---|---|
| Context capacity | Virtually unlimited | Limited by model |
| Token usage | Higher (coordination overhead) | Lower |
| Speed | Slower (subagent coordination) | Faster |
| Task complexity | Handles complex multi-part work | Best for focused tasks |
For trivial tasks, the coordination overhead isn’t worth it. For ambitious work, agents unlock capabilities that would otherwise be impossible.
What’s Next
Section titled “What’s Next”- Enabling Agents - How to toggle agents mode before starting a task
- Monitoring Subagents - Track subagent activity and progress
- Parallel Execution - How Multi runs subagents and tool calls concurrently
- Handling Actions - Respond to subagent requests and approvals
- Usage Tracking - Understand token consumption across your agent hierarchy