Module 6.2 · Topic 2
Multi-Agent Patterns
Bottom Line Up Front: When a single agent cannot complete a task effectively alone, multiple specialized agents coordinating through defined patterns accelerate work and improve quality. Understanding when and how to…
2.1 When One Agent Isn't Enough
Recognizing the limits of a single agent helps you design effective multi-agent solutions. Some situations clearly call for multiple agents.
Trigger 1: Sequential Expertise
The task requires different expertise at each stage. Example: A legal contract review workflow needs a researcher to identify precedents, a language expert to analyze clause nuance, and a risk analyst to flag exposure. One agent attempting all three roles dilutes quality. Assigning each to a specialist agent and routing the contract sequentially through them produces better outcomes.
Trigger 2: Parallel Speed
Multiple independent subtasks can execute simultaneously. Example: Analyzing a company's financial health requires simultaneously researching market conditions, evaluating competitor positioning, and assessing internal capacity. A single agent does these serially (days). Three agents working in parallel deliver results in hours, then hand off to a synthesis agent.
Trigger 3: High-Stakes Verification
A second agent verifying the first agent's work catches errors. Example: An AI generating investment recommendations should have a second agent fact-check the recommendations and a third verify compliance with regulatory constraints. Each agent specializes in one verification dimension.
Trigger 4: Knowledge Domain Isolation
One agent should not have access to all information. Example: A billing agent shouldn't see client confidential information; a contract agent shouldn't see pricing. Separate agents with separate access controls ensure data flows appropriately while preventing overreach.
2.2 Manager-Worker and Delegation Patterns
The manager-worker pattern is the foundation of multi-agent coordination. A manager agent decomposes a large task, delegates to specialist worker agents, and assembles their results into a unified output.
- Task Analysis and Decomposition: The manager receives a complex task and breaks it into subtasks. "Analyze this enterprise's operational efficiency" decomposes into subtasks: "Analyze financial performance metrics," "Evaluate supply chain operations," "Assess workforce productivity." Each subtask maps to a worker agent.
- Worker Assignment and Delegation: The manager assigns each subtask to a specialist worker agent with appropriate access (data, tools, knowledge base). A finance worker gets budget and revenue data; an operations worker gets supply chain logs. Each worker understands its scope and constraints.
- Parallel or Sequential Execution: Workers execute either in parallel (if tasks are independent) or sequentially (if task B depends on task A's output). The manager coordinates the sequence and provides any intermediate results one worker needs from another.
- Result Collection and Quality Check: As workers complete their subtasks, the manager collects results and validates quality. Has a worker delivered what was requested? Is the output in the expected format? Do results from different workers align or contradict?
- Assembly and Synthesis: The manager integrates worker outputs into a cohesive response. If workers found conflicting information, the manager resolves the conflict through additional query or explicit documentation of the disagreement. The final output reads as a unified analysis, not as disconnected pieces.
Unlock the Full AI Skill Building Experience
Visit LawQi for access to the full AI Skill Building experience.
Visit LawQi40% discount using code REASONABLE for personal subscriptions.
Free 48-hour preview access when investigating for teams.