Pick the Right Topology Before Adding Agents
Most Multi-Agent projects stall not because the agents aren't smart enough, but because they were bolted together before anyone decided what shape they should collaborate in. The collaboration topology governs how tasks flow, who makes decisions, and where things converge on failure; choose the wrong one and more agents just buy you more expensive chaos. This isn't an abstract aesthetic question either—Anthropic's production data and Google Research's controlled study show that running the same task under different topologies can swing token cost and reliability by more than tenfold. This article is strictly about choosing the architectural shape: how agents connect and how decision authority is distributed. It does not cover how they message each other, share memory, or split roles. Get the shape right first, and the implementation has something to stand on.
Four Topologies Mapped to Real Frameworks
| Topology | Representative frameworks | Best fit and real cases | Key trade-off (with data) |
|---|
| Orchestrator-Worker | LangGraph Supervisor, Anthropic Research, CrewAI Hierarchical | Parallelizable exploration: Anthropic Research spins up 3–5 subagents at once | Reliable but costly: Anthropic's version beats a single agent by 90%, at ~15x the tokens |
| Hierarchical | LangGraph multi-level Supervisor, CrewAI Hierarchical | Large tasks whose subtasks can themselves be split further | Scales deep; the manager's extra LLM calls add token cost |
| Peer / Network | OpenAI Agents SDK handoffs, AutoGen Group Chat | Fuzzy-boundary, exploratory work needing on-the-fly handoffs | Most flexible, most fragile: error amplified up to 17.2x without a center |
| Pipeline | CrewAI Sequential, MetaGPT SOP | Fixed linear flows, e.g. PRD→architecture→code | Lowest token cost, easiest to debug; one stalled stage halts the line |
Orchestrator-Worker: The Data from Anthropic and LangGraph
Orchestrator-worker is the most proven starting point: one orchestrator agent decomposes and delegates, while workers each handle a slice and report back. Anthropic's Research system is exactly this shape—the lead agent plans, spins up 3–5 specialized subagents to gather evidence in parallel, then synthesizes and runs a separate citation pass. In their tests, a Claude Opus orchestrator with Sonnet subagents outperformed a single Opus by 90.2%, and token usage alone explained about 80% of the performance variance. The price: it burns roughly 15x the tokens of a plain chat. On the engineering side, LangGraph's langgraph-supervisor standardizes the pattern: a supervisor node uses Command to decide which worker node runs next—clear ownership and easy tracing, but the supervisor is inherently a bottleneck and single point of failure.
Hierarchical: Add Layers Only When Scale Demands It
When subtasks themselves grow large enough to need splitting, promote orchestrator-worker to a hierarchy: the top orchestrator does coarse division and hands each block to a sub-orchestrator for finer breakdown. LangGraph directly supports "a supervisor that manages multiple supervisors" to build multi-level teams; CrewAI's Hierarchical Process automatically or explicitly assigns a manager agent that dynamically delegates, reviews output, and orders rework. This scales beyond what a single layer can carry, but layers cost money—the manager itself consumes LLM calls to analyze tasks, pick agents, and review output, so every added layer brings extra coordination overhead and token cost. The rule is blunt: only add a layer when a single supervisor's delegation list grows too long for it to track. Don't stack layers preemptively for tidiness.
Peer: Flexibility That Buys a 17x Error Amplification
Peer topologies have no central command—agents are equals, and whoever is capable takes the next step. OpenAI evolved its early Swarm into the Agents SDK, whose core is handoffs: an agent facing a request outside its domain passes control, along with context, to a better-suited peer. AutoGen's Group Chat offers auto and round_robin strategies for picking the next speaker. Peer suits fuzzy-boundary exploratory work where steps can't be planned ahead—maximally flexible, yet hardest to predict and debug. Google Research's controlled study "Towards a Science of Scaling Agent Systems" (December 2025) measured the brutal number across 180 controlled configurations: an independent multi-agent setup with no central coordination amplifies errors to 17.2x a single agent, while adding a coordinator compresses that to 4.4x. So peer fits small, high-autonomy groups that tolerate trial and error—not critical flows that demand stable output.
Pipeline: Cheapest and Most Traceable, but Brittle
Pipeline is the opposite: tasks flow through stations in order, each focused on one operation, passing its refined output to the next. CrewAI's Sequential Process is the textbook case—tasks run in defined order, and each step's output automatically becomes the next step's context. With no coordination overhead, it has the lowest token cost and is the easiest to debug. MetaGPT takes the assembly line to the extreme with its "Code = SOP(Team)" philosophy, chaining product manager, architect, project manager, and engineer into one SOP line that turns a single requirement into PRD, design, and code. Pipelines are most efficient and observable when steps are fixed, and easiest to insert checkpoints between; but they are inherently fragile—one stalled station halts the whole line, so every station needs explicit timeout, retry, and failure handling.
Nerdtechnic: Deriving Architecture from Task Characteristics
As the data shows, topology selection isn't about which pattern is newest—it's about reverse-engineering from a task's decomposability, scale, and flow shape, while trading off reliability against token cost: pay the centralization overhead for stability, go pipeline to save, and reach for peer only when you need flexibility. The pragmatic path for most enterprises is to start with orchestrator-worker, grow into a hierarchy when scale demands it, distill into a pipeline when the flow is fixed, and mix as tasks evolve. Nerdtechnic's AI systems consulting first maps your actual task structure, then matches it to the most suitable collaboration topology and framework choice—so the system grows in the right shape from day one, instead of buckling in production and forcing a rebuild.
References
- Anthropic, "How we built our multi-agent research system", 2025. Source
- Google Research, "Towards a science of scaling agent systems: When and why agent systems work?", 2026. Source
- LangChain, "langgraph-supervisor-py (GitHub)", 2025. Source
- OpenAI, "Swarm (GitHub; notes the Agents SDK as its production successor)", 2025. Source
- Microsoft, "AutoGen 0.2: agentchat.groupchat reference", 2024. Source
- FoundationAgents, "MetaGPT: The Multi-Agent Framework (GitHub)", 2025. Source
- CrewAI, "Sequential Processes (official docs)", 2026. Source
- CrewAI, "Hierarchical Process (official docs)", 2026. Source
Technical Sharing
