Coordination patterns are useful frameworks, but what does a working agent team actually look like? Here are five configurations we've seen produce consistently strong results, along with the reasoning behind each design choice.
Problem: "Analyze our top 5 competitors in the project management space"
Why Fork-Join: Each competitor can be analyzed independently (the fork), then all findings merge into a comparative matrix (the join). This maximizes parallelism while ensuring a unified output.
Team (5 agents):
Why it works: Competitor research is embarrassingly parallel — each company's data is independent. The synthesis step is where the real insight happens, comparing across all profiles to find patterns no single analysis would reveal.
Problem: "Create a 4-week content marketing campaign for our new API product"
Why Sequential Pipeline: Content strategy flows naturally through phases — audience research informs messaging, messaging informs content creation, and content creation informs the distribution plan.
Team (4 agents):
Why it works: Each phase builds directly on the previous one. The audience researcher's output literally shapes what the content creator writes. Skipping or parallelizing these steps would produce generic content disconnected from the audience.
Problem: "Evaluate whether we should acquire this B2B SaaS company"
Why Advisory Debate: Acquisition decisions are high-stakes and benefit from adversarial analysis. You want someone arguing for the deal AND someone poking holes in it.
Team (4 agents):
Why it works: Confirmation bias is the enemy of good investment decisions. By structurally requiring a negative case, you surface risks that an optimistic single analysis would miss. The synthesizer's job is to find truth, not consensus.
Problem: "Plan the launch of our mobile app in 3 new European markets"
Why Parallel Workers: Each market requires similar analysis but with completely different data. Germany, France, and Spain each need localization research, regulatory review, and go-to-market planning — all independent work streams.
Team (6 agents):
Why it works: The per-country research is completely independent and runs simultaneously. The coordination agents (localization, launch, risk) only activate after the parallel phase completes, ensuring they work with complete information.
Problem: "Analyze 6 months of customer feedback to find product improvement opportunities"
Why Subagent Scout: Customer feedback comes from many sources — support tickets, NPS surveys, app reviews, social media, sales call notes. A scout agent can dispatch specialists to each source, then synthesize the findings.
Team (5 agents):
Why it works: Each feedback source requires different analytical approaches. Support tickets need issue categorization; reviews need sentiment analysis; surveys need statistical interpretation. Specialists produce better analysis than a generalist trying to cover all sources.
These examples share a common principle: the team structure mirrors the problem structure. Independent data sources get parallel agents. Sequential dependencies get pipeline agents. High-stakes decisions get debate agents. Match the pattern to how the work naturally decomposes, and your agent team will produce results that feel coordinated rather than stitched together.