Among the coordination patterns available for AI agent teams, Supervisor-Worker is the most commonly used — and the most commonly misused. It's the pattern people reach for by default, but getting it right requires understanding when centralized coordination actually helps versus when it just adds overhead.
The Supervisor-Worker pattern has one agent (the Supervisor) that breaks down the problem, assigns tasks to specialist Worker agents, reviews their output, and synthesizes the final result. The Supervisor doesn't do the actual work — it orchestrates.
Think of it like a senior manager who understands the big picture, delegates effectively, and quality-checks the results. The workers are specialists who go deep on their assigned tasks.
In a Sequential Pipeline, the work is pre-defined: Agent A does step 1, Agent B does step 2. But what if step 1's output reveals that step 2 needs to be completely different than expected?
The Supervisor adapts. After reviewing Worker A's output, it can reassign Worker B to a different task, spawn additional work, or skip unnecessary steps. This dynamic allocation is impossible in rigid pipeline or parallel patterns.
The Supervisor acts as a quality gate between phases. If Worker A produces subpar analysis, the Supervisor can request a revision before Worker B builds on that foundation. This prevents errors from compounding through the pipeline.
Some projects have dependency graphs that don't fit neatly into sequential or parallel patterns. Task C might depend on Tasks A and B, while Task D only depends on A, and Task E depends on C and D. The Supervisor manages these complex dependencies naturally, dispatching work as prerequisites complete.
The Supervisor agent needs three things in its prompt:
The Supervisor must understand the domain well enough to decompose the problem intelligently. It doesn't need to be an expert in every worker's specialty, but it needs to know what specialties exist and how they connect.
Example: For a market entry analysis, the Supervisor should know that regulatory review should happen before go-to-market planning (because regulations might make certain strategies impossible), even though competitive analysis can happen in parallel with both.
The Supervisor needs to know what workers are available and what each one does. Without this, it can't delegate effectively.
Strong roster specification:
You have 4 workers available:
- Financial Analyst: Builds financial models, runs projections, calculates ROI
- Market Researcher: Gathers market data, sizes markets, identifies trends
- Legal Reviewer: Assesses regulatory requirements, compliance risks, IP concerns
- Strategy Writer: Produces executive-ready documents, presentations, and recommendations
The Supervisor needs to know what "done" looks like. Without clear success criteria, it might over-optimize one area while neglecting another.
Example: "The final deliverable is a 10-page market entry recommendation that covers market opportunity (with sizing), competitive landscape (minimum 5 competitors), regulatory requirements (by country), financial projections (3-year horizon), and go/no-go recommendation with confidence level."
The worst Supervisor-Worker implementations have the Supervisor micromanaging every decision. If the Supervisor is rewriting worker output instead of reviewing it, you've created a bottleneck. Give workers autonomy within their domain and have the Supervisor focus on coordination, not execution.
"Research the market" is not a useful task assignment. "Identify the top 5 competitors in the European B2B project management space, analyze their pricing tiers, and estimate their annual revenue" gives the worker a clear, completable task.
The Supervisor's most important job happens last — synthesizing all worker outputs into a coherent whole. If you skip this step, you get a collection of good individual pieces that don't tell a unified story.
Use Supervisor-Worker when: Tasks have complex dependencies, quality gating matters, or the work plan might change based on intermediate results.
Use Pipeline when: The work has a clear linear sequence and each step's requirements are known upfront.
Use Parallel Workers when: All tasks are independent and no coordination is needed during execution.
Use Advisory Debate when: The goal is to stress-test a specific decision rather than produce a broad deliverable.
Supervisor-Worker is the Swiss Army knife of coordination patterns. It handles complexity, adapts to surprises, and maintains quality through centralized oversight. But like any Swiss Army knife, it's not always the best tool for a specific job. Use it when you genuinely need dynamic coordination — and reach for simpler patterns when you don't.