Cargando el Technology Radar
Multi-agent systems coordinate specialized agents to handle complex workflows through delegation, routing, handoffs, subagents, or custom orchestration. LangChain describes multi-agent systems as coordinating specialized components for complex workflows, while warning that not every complex task requires multiple agents (LangChain Docs).
The strongest reason to use multi-agent designs is context engineering. Anthropic describes sub-agent architectures as a way to handle focused tasks with clean context windows, where subagents explore deeply and return condensed summaries while the lead agent synthesizes results (Anthropic Engineering).
Keep this in Assess because the pattern is powerful but easy to overuse. Multi-agent systems should be justified by measurable gains in parallelism, context isolation, specialization, or review quality, not by architectural novelty.
Coordination overhead can exceed benefits. LangChain’s performance comparisons show multi-agent patterns can add model calls, tokens, and latency, especially when sequential handoffs or repeated stateless subagent calls are used (LangChain Docs).
Failure modes compound across agents. One agent’s bad summary, unsafe tool output, or hallucinated intermediate result can become trusted input for another agent unless outputs are validated and provenance is preserved.
Authority boundaries are hard. Teams need clear rules for which agent can call which tools, which agent owns the final answer, when humans approve actions, and how conflicts between agents are resolved.
Security and cost controls become harder. More agents means more prompts, more tools, more context copies, more traces, and more places where prompt injection, data leakage, or runaway loops can occur.
Assess multi-agent designs only when decomposition creates measurable value: parallel research, independent review, specialized domain context, large tool surfaces, or long-horizon workflows. Start with the simplest pattern that works, such as a router or planner-executor, before adding autonomous subagent networks.
Require orchestration controls: task boundaries, budgets, tool permissions, traceability, summarization rules, provenance, failure handling, and a clear owner for final decisions. For many enterprise workflows, one well-instrumented agent with dynamic tools and strong evals is easier to operate.