Multi-Agent
When Do You Need Multiple Agents
More Agents isn't always better. Most of the time, one is enough. But there are three real scenarios where a single Agent can't handle it alone. Know when you actually need to divide the work.
Three Real-World Scenarios
Parallel Acceleration
Search 5 sources simultaneously — 5x faster than sequential search
Real Case: News Digest
User requests a summary of today's AI news. A single Agent must search 5 websites sequentially, taking 25 seconds total.
Replace with 5 sub-Agents searching simultaneously, each covering one source; the main Agent aggregates the results — total time: 5 seconds.
Key insight: These searches are independent of each other, making them naturally parallelizable.
User requests a summary of today's AI news. A single Agent must search 5 websites sequentially, taking 25 seconds total.
Replace with 5 sub-Agents searching simultaneously, each covering one source; the main Agent aggregates the results — total time: 5 seconds.
Key insight: These searches are independent of each other, making them naturally parallelizable.
Role Separation
One writes code, another reviews it — mutual checks and balances
Real Case: Code Review
Asking the same Agent to write code and then review its own code rarely surfaces errors — much like proofreading your own essay.
Split into two Agents: Writer writes the code, Reviewer reviews it. The Reviewer doesn't know Writer's reasoning — only sees the final code — making it far more effective at catching issues.
Key insight: Role isolation makes the review genuinely effective.
Asking the same Agent to write code and then review its own code rarely surfaces errors — much like proofreading your own essay.
Split into two Agents: Writer writes the code, Reviewer reviews it. The Reviewer doesn't know Writer's reasoning — only sees the final code — making it far more effective at catching issues.
Key insight: Role isolation makes the review genuinely effective.
Risk Isolation
Sub-task failures don't cascade to the main task
Real Case: Complex Document Processing
The main Agent is compiling a 50-page report that requires parsing several PDF attachments.
If PDF parsing fails (corrupted format, timeout), handling it directly in the main Agent will crash the entire task.
Delegate to a sub-Agent to handle the PDF separately: if it succeeds, report the result; if it fails, report "this file has an issue" — the main Agent continues unaffected.
Key insight: Sub-task failures are contained.
The main Agent is compiling a 50-page report that requires parsing several PDF attachments.
If PDF parsing fails (corrupted format, timeout), handling it directly in the main Agent will crash the entire task.
Delegate to a sub-Agent to handle the PDF separately: if it succeeds, report the result; if it fails, report "this file has an issue" — the main Agent continues unaffected.
Key insight: Sub-task failures are contained.
Parallel Acceleration Flow Diagram
One task split into multiple sub-Agents running in parallel
Main Agent
Dispatch tasks
Dispatch tasks
→
Run concurrently
Search A
Search B
Search C
→
Aggregate Results
Decision Criteria: Before adding a second Agent, ask yourself:
① Can't a single Agent really do this? (Often it's just a poorly written Prompt)
② Is the added complexity worth it? (More Agents means more coordination cost and more failure points)
③ Is there a simpler solution? (e.g. parallel tool calls might be enough without actually splitting into Agents)
① Can't a single Agent really do this? (Often it's just a poorly written Prompt)
② Is the added complexity worth it? (More Agents means more coordination cost and more failure points)
③ Is there a simpler solution? (e.g. parallel tool calls might be enough without actually splitting into Agents)
More Agents isn't always better — most of the time, one is enough. Know when you actually need to divide the work. Multi-Agent is only worth introducing when there's a clear need: parallel acceleration, role checks, or risk isolation.