Multi-Agent Systems

A multi-agent system runs multiple LLM-backed agents — typically with distinct roles, tools, prompts, or model choices — coordinating on one task. The motivation is the same as in human teams: specialization beats generalism on complex work, and decomposition makes each piece tractable.

Common patterns

• Coordinator + specialists. A top-level agent plans and dispatches sub-tasks to specialist agents (a coding specialist, a research specialist, a data-analysis specialist). The coordinator integrates results. Cleanest, most-shipped pattern in production.
• Sequential pipeline. Agents arranged in a chain — one drafts, the next reviews, the next polishes. Conceptually similar to a feature pipeline; popular in content workflows.
• Debate / adversarial. Two or more agents argue different sides of a question; a judge agent (or majority vote) decides. Useful for hard reasoning problems where errors are uncorrelated across agents.
• Marketplace / blackboard. Agents post partial solutions to a shared workspace; other agents read and contribute. Powerful in theory; messy in practice for debugging and bounding cost.
• Hierarchical decomposition. Coordinators of coordinators: the top-level planner decomposes into sub-coordinators, who decompose further. Scales to large goals; risks explosive cost without aggressive budgets.

Why specialization actually wins

The naive assumption is that multi-agent benefits come from “cognitive division of labor” — like a human team where each person thinks differently. That’s mostly not what happens with LLMs (especially same-family LLMs). The real wins:

• Context discipline. Each specialist has a focused system prompt and a focused tool subset. Less context means less attention dilution and more reliable tool selection .
• Per-agent model choice. A frontier LLM for the planner; cheaper specialized models for narrow sub-tasks. Cost falls dramatically without quality loss.
• Independent reasoning paths. In debate or multi-sample setups, having different agents (different models, different prompts) gives uncorrelated error patterns. Aggregate beats the best individual.

Where it breaks

• Cost multiplication. N agents per turn = N× token spend, plus inter-agent message overhead. A 3-agent debate plus reflection is easily 8× a single-agent baseline.
• Latency. Sequential coordination is slow; even parallel multi-agent has aggregation latency.
• Coordination errors. The coordinator gets the sub-task description subtly wrong, the specialist solves the wrong problem, and the coordinator integrates the wrong-problem solution into the final answer. Hard to debug because each agent looks individually correct.
• Same-model collusion. Multiple agents drawn from the same model often share blind spots — they confidently agree on the same wrong answer.

The pragmatic stance

Most production multi-agent systems are “coordinator + 2-4 specialists, each with a focused tool set.” Debate, marketplace, and deep hierarchies are research-grade patterns that ship occasionally. What pays off: each agent gets a tight role, a focused context, and the cheapest model that can do its job. Frontier LLM as the coordinator; specialized models — including rerankers and context compressors — wherever the sub-task is narrow enough to swap in.