Pick the honest default architecture for a brand new agent feature in 2026

MCQ·Easy·4.0 · 0·~1 min·Asked atDataiku Perplexity Wipro

Attempt it

The 2026 production consensus is clear: **start single-agent**. A single well-prompted agent with the right tools is faster, cheaper, more debuggable, and usually higher quality than the multi-agent alternatives for any workload that fits in one system prompt. Reach for multi-agent only when a **specific named failure of the single-agent version** shows up. Multi-agent does buy real things: parallel exploration over independent branches, role-shaped sub-prompts, bounded tool subsets per agent, and structurally enforced human in the loop checkpoints. But it pays for these in coordination cost (one supervisor LLM call per hop), latency (calls are serial), cost per task (often 5-10x the single-agent equivalent), and debuggability (the trace becomes a graph). The other three options in the question are anti-patterns when picked as defaults: - **Debate council**: voting adds 3x cost for marginal quality gains on most tasks. - **Supervisor plus three role-played workers**: separation of concerns is not free; it costs an LLM call per delegation. - **Peer swarm**: handoff loops and ping-pong failures are easy to introduce; debugging is hard. The industry pattern: Devin, OpenHands, Claude Code subagents, Cursor agent mode all ship a strong **primary loop** (single agent doing most of the work) and spawn subagents only for well-scoped sub-tasks where the split is justified. The default is one; the escalation is conditional.

**Start single-agent.** The 2026 production consensus and the right answer for most 'should I use multi-agent' questions. A single well-prompted agent with the right tool belt handles most workloads at a fraction of the cost, latency, and operational complexity of multi-agent alternatives. Multi-agent earns its overhead only when one of these failure modes is real: - Context window saturates before task completion (planner-executor helps). - Tool surface exceeds ~15-20 and reliability degrades (supervisor plus workers helps). - Independent sub-queries can run in parallel (parallel research plus reducer helps). - Verifiable quality bar wants iteration (critic loop helps). - Structurally enforceable HIL checkpoint required (multi-agent boundary helps). If none apply, multi-agent pays coordination overhead for no structural gain. A two-agent run costs 3-5x the single-agent equivalent. Chatty crews hit 10x. Latency stacks linearly. Debuggability collapses from linear to graph trace. **The anti-defaults**: - **Debate council**: 3x cost, marginal quality gain on most tasks. Reserve for adversarial fact-checking. - **Supervisor plus three workers**: 'separation of concerns' costs an LLM call per delegation and needs a real routing decision to justify. - **Peer swarm**: handoff loops are easy, traces are graphs, ping-pong bites production hard. Anthropic's published guidance, Devin, Cursor, and Claude Code subagents all converge on the same pattern: strong single-agent primary loop, conditional subagent spawning only when the workload genuinely demands it.

TL;DR

Start single-agent in 2026; escalate to multi-agent only when a specific single-agent failure justifies the cost and complexity.

Memory aid

Easy to grasp

Imagine being asked to start a small kitchen. You could hire one excellent cook with a good set of knives and pans, or you could hire three specialists - a sous chef, a saucier, and a pastry chef - and a head chef to coordinate them. The three-cook setup looks impressive, but it costs more, every dish has to wait for the head chef to assign it, and a mistake in one station can ripple across the others. Most days, one excellent cook with a good toolkit can make everything you need. You hire the second and third cooks when the single cook genuinely cannot keep up or specifically cannot do a dish - not because three cooks sound fancier than one.

Key concepts

Multi-Agent Systems

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

The 2026 production consensus on multi-agent system design is clearer than the 2023-2024 hype suggested: start single-agent. A single well-prompted agent with the right tool belt is the right default for most production agent features. Multi-agent earns its overhead only when the workload structurally demands it, and 'structurally demands' is a high bar with specific named failure modes attached.

This question is essentially asking whether you understand that consensus. The other three options - debate council, supervisor with three role-played workers, peer swarm - are not bad patterns in absolute terms, but they are wrong defaults. They have specific niches; none of them is the right answer for a brand-new agent feature in 2026.

This walkthrough covers why single-agent wins as default, the named-failure bar for escalating to multi-agent, what the industry actually ships, and the migration discipline that separates production-sustainable agent systems from architecturally impressive ones that burn budget.

Mental model: the default is one. The escalation is conditional. The condition is a named, observable failure of the single-agent version - never aesthetics.

Why single-agent wins by default in 2026

Cost

A single-agent turn is one LLM call. A two-agent supervisor-worker turn is three LLM calls (supervisor routes, worker executes, supervisor reviews). A chatty 5-agent crew at 8 turns can hit 40-80 LLM calls per task. At fleet scale and frontier-model rates, the cost difference between single and multi can be 10-50x.

The most common reason production teams rolled back from multi-agent in 2024-2025 was cost. Quality was often acceptable; cost was the blocker.

Latency

LLM calls are mostly serial. A two-agent split doubles the wall-clock time. A supervisor plus workers crew triples or quadruples it. For interactive UX with sub-second budgets, multi-agent often blows the budget before quality matters.

Debuggability

A single-agent trace is linear. A multi-agent trace is a graph. Post-incident review on a graph trace is materially harder: which agent made the bad decision, did the handoff payload look right, did the supervisor route correctly. Linear traces have one path to read; graph traces have many.

Failure modes

Single-agent fails in ways teams understand: bad tool calls, context overflow, format violations. The remediation patterns are well-established.

Multi-agent fails in new ways: handoff loops, role bleed, deadlocks, supervisor bottlenecks, ping-pong. Each new failure mode requires new instrumentation and new mitigation patterns. Teams new to multi-agent often experience these failures in production before they have observability to debug them.

Operational surface

One prompt vs N prompts to version and roll out. One eval set vs N to keep aligned. One on-call runbook vs N. The ops cost doubles in pieces for every agent added.

Quality on fit workloads

For workloads that fit single-agent shape (most production features), a well-tuned single agent often beats a poorly-tuned multi-agent system. The investment in tuning one agent goes further than splitting investment across many. This is the inverse of the common assumption that multi-agent is automatically higher quality.

The named-failure bar

What the industry actually ships

Why the other three options are anti-defaults

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

Cursor agent mode ships as a single agent with a large tool belt; the team chose single-agent for debuggability and latency over multi-agent alternatives.
Claude Code uses a strong primary single-agent loop and spawns subagents only for well-scoped sub-tasks; most coding sessions never spawn a subagent.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhat signal would push you from single-agent to a planner-executor split?

Trace analysis showing context window saturating before task completion, plans being forgotten or contradicted mid-execution, or token spend dominated by re-reading earlier tool outputs. Quantitative trace signal beats aesthetic preference.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Starting multi-agent because it sounds more sophisticated. The architectural sophistication earns its keep only when the workload structurally demands it; otherwise it is cost and complexity for no quality gain.

60 second bullets to scan on the way to the call.

Why single-agent is the production default in 2026
The specific named failures that justify multi-agent

Primary sources. Browse if you want the original framing.