
A taxonomy of multi-agent coordination and lessons for manufacturing
An AWS post presents multi-agent coordination as a question of workflow and control. Its taxonomy separates two levels. Tier 1 contains deterministic control flows whose paths are established in advance. In Tier 2, the model selects paths at runtime. The architectural guidance is to begin with a direct model call, use Tier 1 when the solution path is known, and move to Tier 2 only when exploration discovers subtasks during execution.
Deterministic coordination patterns
A sequential pipeline is the basic ordered pattern. Its nodes are arranged at build time, and every node passes its output to the next one. This makes the progression explicit before execution. A conditional workflow adds logic-based branches. It can also impose bounded retries and escalate to a human after the permitted attempts are exhausted.
A conditional router performs a narrower operation: it classifies each input and runs exactly one processing path for each request. The branches are alternatives rather than a collection of paths that all execute together.
A parallel workflow applies when subtasks are independent. Those subtasks run concurrently, and a fan-in step combines their results. If one path fails, only that path is retried. Its latency is governed by the slowest parallel agent, not by the sum of the running times of every agent. The distinction matters when comparing parallel execution with a sequential pipeline: one accumulates ordered stages, while the other waits for its longest concurrent branch before combining results.
Runtime-selected paths
Plan-based orchestration gives a central orchestrator an explicit plan. The orchestrator assigns slices of work to stateless, independent workers, monitors their progress, and retries only a worker that fails. The explicit plan and the worker assignments provide the organizing structure.
The handoff pattern uses local, peer-to-peer delegation instead of relying only on a central assignment mechanism. When control moves, agents transfer validated task state, supporting evidence, and unresolved questions. The handoff therefore covers both control and the information needed to continue the task.
Conversation-driven patterns place all participating agents in a shared conversation. Every agent observes the messages, while state resides implicitly in conversation history. Turn-taking may follow a round-robin order or be selected by the model. This differs from an explicit central plan because the shared history carries the working state.
Debate is a more expensive pattern. Multiple agents independently assess the same input before their conclusions are reconciled. Reconciliation may use a majority, a judge agent, or escalation. The post reserves this approach for high-stakes, low-volume decisions, rather than treating it as a default for routine work.
Anti-patterns and architecture choices
The post identifies several failure-prone designs: agents equipped with too many tools, retry loops with no bound, prose handoffs without a schema, multi-agent systems where multiple agents are unnecessary, and real-world actions performed without approval. Together, these warnings reinforce the selection rule in the taxonomy: additional coordination should be introduced only when the work requires it.
The baseline is a direct model call. If the path is already known, deterministic Tier 1 patterns provide the relevant structure. Tier 2 is intended for work in which exploration itself reveals new subtasks. The dividing line is therefore not simply whether several components can be created, but whether the execution genuinely needs runtime discovery and model-selected routing.
Manufacturing guidance
For manufacturing safety, the post recommends keeping PLC, SCADA, and MES control deterministic. Agents can observe, analyze, and recommend, while deterministic systems retain control. Safety constraints should exist outside the agents as deterministic guardrails. Before actuation, a result should be validated and receive approval at a level appropriate to the proposed action.
The operating model also combines edge and cloud roles. Edge systems handle time-sensitive local analysis. Cloud systems handle broader analytics and retraining. The guidance pairs this hybrid arrangement with shift-bounded agent memory: each shift ends with a structured handoff to the next crew. This creates an explicit transition between crews while preserving the distinct roles of local analysis, broader analytics, and deterministic control.
The taxonomy consequently offers a progression rather than a requirement to maximize the number of agents. Start with the direct call, add deterministic workflow structure when the route is understood, and use runtime-selected coordination only when subtasks emerge through exploration. The complete organization-specific discussion appears in the original AWS post.