Which conditions favour a multi-agent architecture over a single-agent system?
Multi-agent wins only when subtasks parallelise, specialist models beat a generalist, or volume overflows one context. Otherwise it just adds coordination cost.
Imagine you have a big project due. If the work splits into pieces that people can do at the same time without waiting on each other, hiring a team helps a lot. If one teammate is great at math and another at writing, splitting by skill also helps. But if the work is one long chain where step two needs step one finished first, a team does not finish faster. You just add the cost of everyone talking to each other and handing things off. AI agents are the same. A single agent is the simple default. You only split into a multi-agent team when the task genuinely parallelises, when specialists clearly beat a generalist, or when one agent simply cannot hold all the work in its head at once.
Detailed answer & concept explanation~7 min readEverything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Everything important, quickly.
State that single-agent is the default, then name the three conditions that justify multi-agent: parallel independent subtasks, specialist advantage with context isolation, and scale beyond one context window. Explain why a sequential chain and 'simpler orchestration' are distractors, and close with coordination overhead and error propagation as the costs you weigh against the benefit.
| Dimension | Single-agent | Multi-agent |
|---|---|---|
| Latency on parallel work | Serial, slow | Fan-out, fast |
| Coordination overhead | None | High: routing, handoffs, merge |
| Error propagation | Contained to one loop | Compounds across agents |
| Context isolation | Single shared context | Clean per-agent contexts |
| Default choice | Yes, the baseline | Only when a real benefit wins |
Real products, models, and research that use this idea.
- Anthropic's multi-agent research system uses an orchestrator that spawns parallel subagents to explore independent search branches, reporting large latency wins over a single agent on broad research tasks.
- OpenAI's Swarm and the newer Agents SDK model handoffs where a triage agent routes to specialist agents, a specialisation pattern rather than a parallelism one.
- LangGraph supervisor architectures coordinate worker agents over a shared typed state, exposing the coordination overhead and termination logic that single-agent loops avoid.
- CrewAI assigns role-specialised agents (researcher, writer, reviewer) to a shared goal, useful only when the role split genuinely beats one generalist agent like Claude Opus 4.7.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you decide the boundary between subagents in a research task, and what makes a clean split?
QWhy does a multi-agent system often cost more total tokens than a single agent on the same task?
Don't say thisRed flags and common mistakes that signal junior thinking. Click to expand.
Red flags and common mistakes that signal junior thinking. Click to expand.
Reaching for a multi-agent design by default. A team of agents adds coordination cost and error propagation, so a single agent should be the baseline you must justify moving away from.
The night-before-the-interview bullets. Scan these on the way to the call.
Primary sources. Skim if you want the original framing.
Same topic, related formats. Practice these next.