Malaysia

All field notes

One agent or a team?

October 24, 2026 · By

Context: Our own experiment · domain: AI agents / multi-agent design

The problem

A small business buying an AI agent soon hits a familiar question: one agent, or a team of agents — a lead to plan the work, specialists to run it, and a critic to check the answer?

The real question is simpler: on my job, does a team beat one agent, and what does it cost in time and tokens, the units AI models bill by?

Our test was inspired by a 23 October 2026 article by 码良 (@cxjwin), “Agent 高频题 03. 为什么要多 Agent:怎么拆、怎么通信、结果谁汇总”. It asked whether anyone had compared a multi-agent team against a single-agent baseline. The author noted that their own run (1 lead plus 3 subagents, 62 minutes, about 192 times its token budget) had no such baseline.

We built that baseline and ran both side by side.

What people usually try

Splitting a job into a lead and specialist workers looks organised and promises parallel speed.

In practice, coordination costs time and tokens. The lead writes briefs. Each worker reads its documents and writes a report. The lead waits for all of them and merges the reports, and a critic reads the merged answer. Even when workers run in parallel, briefing, merging and checking run one after another, and every handoff re-reads reports.

Others report a similar pattern. In Anthropic’s 22 October 2026 note, “Building agents: When to fold subagents into your main agent loop”, Diane Sarkis reported that in several enterprise deployments a single agent with skills often beat a subagent design on quality, and often at lower cost and latency. One customer saw roughly 2× faster runs on a single loop. Subagents still earned their place for verification, parallel work and separate permissions.

Kim et al. (“Towards a Science of Scaling Agent Systems”; blog) compared single-agent and multi-agent setups across 260 configurations, holding the model, tools and prompts the same within each comparison. Multi-agent setups did about +81% on parallelisable financial reasoning and −39% to −70% on sequential planning. Returns also shrank once one agent was already strong.

What we found

Across 24 measured runs (3 runs per setup across two jobs and two models), the team never clearly beat one agent on quality, and it took 2× to 11× as long and 1.5× to 24× the tokens — about 2× as long on the competitor check and 5–11× as long on the treatment quote.

We set the test in “Juniper Family Dental”, a fictional four-chair US dental practice. It is not a real business or a client, and the five competitor practices are fictional too. The two jobs drew on 31 fixed local documents with no live web, so every run could be scored against a gold answer.

One agent used a single loop with all the tools. The team used a lead, workers returning short fixed-shape reports, the lead merging them, and a critic. Both setups used the same model (deepseek-v4.1-flash or claude-haiku-5-5, with thinking set to high). Each run had one token-and-time budget shared by all its agents, checked before every model call. A run stopped before its final answer scored 0.

We tested two jobs and wrote down a prediction for each before running anything:

  1. Competitor check (parallel shape, 25 documents): compare five fictional practices on six items (new-patient exam price, Saturday hours, online booking, insurance plans accepted, same-day emergencies, review rating) to fill a 30-cell table. The shared budget was 1,200,000 tokens and 600 seconds. The team used 7 agents: a lead, 5 workers and a critic. We predicted a team would help because each worker gets a clean context. Didn’t hold on quality: over 3 runs each on both models, one agent already scored 1.00 (30/30) every time — including the Saturday hours, where an older hours page said “Closed Saturdays” and a newer news post said 9am–2pm. The team had nothing to add and took 2.0–2.1× as long.
  2. Treatment quote (step-by-step shape, 6 documents): calculate what a patient pays for a crown, one filling and a cleaning after insurance and a membership discount across 5 line items (gold total: USD 956.25). The shared budget was 450,000 tokens and 600 seconds. The team used 5 agents: a lead, 3 workers and a critic. We predicted a team would hurt because handoffs drop running numbers. Didn’t hold on quality either: over 3 runs each on deepseek-v4.1-flash, both setups scored 1.00 (5/5), while the team took 5.3× as long and 3.8× the tokens. Deductible, coverage, annual maximum and discount depend on each other, so workers could not work independently. The lead had to settle conflicts over whether preventive care counts against the USD 1,500 annual maximum.

On claude-haiku-5-5, the treatment quote went wrong for both setups. Two of the three one-agent runs scored 0.40 (2/5). Both got the total (USD 956.25) and membership discount (USD 168.75) right, but listed full fees for the cleaning, filling and crown (USD 125, USD 210, USD 1,240) instead of what the patient pays (USD 0, USD 66, USD 1,059). The third one-agent run scored 1.00.

Meanwhile, two of the three team runs on claude-haiku-5-5 scored 1.00. Team run 2 hit the budget at 465,661 of 450,000 tokens and stopped at 189.8 seconds without a final answer, scoring 0. The budget is checked before each model call, so the last calls can run past it. Over 3 runs each, 0.67 vs 0.60 was within run-to-run spread, at 10.8× the time and 24.2× the tokens.

The team compared with one agent, over 3 runs each:

Job · model Quality Time Tokens
Competitor check · DeepSeek Same (all 1.00) 2.1× 2.2×
Competitor check · Haiku Same (all 1.00) 2.0× 1.5×
Treatment quote · DeepSeek Same (all 1.00) 5.3× 3.8×
Treatment quote · Haiku No clear difference (0.67 vs 0.60) 10.8× 24.2×

The lead used 4–33% of the team’s tokens; workers and the critic spent the rest. At list prices, one-agent runs came to USD 0.001–0.006 and team runs to USD 0.016–0.033. On a mean basis, the team cost about 4× on the competitor check, about 6× on the DeepSeek quote, and about 22× on the Haiku quote.

When a team does make sense. Bigger work with genuinely independent parts that each need many steps (what we predicted for the competitor check; on our job one agent already scored 1.00, so the team had nothing to add), a separate checker, or separate permissions.

What it means for a small business buying an AI agent. Start with one well-built agent. Ask any vendor selling “a team of agents” for a measured comparison against one agent on your job — quality, time and cost.

Limits. Two jobs, 3 runs per setup, two low-cost models, and a fictional business with 31 fixed documents. That shows a pattern on these jobs, not a law. Bigger, messier jobs with more independent parts could come out differently.

Watch every run step by step, with each agent’s spend, in the replay.


Looking for a practical setup for your business? See our AI Front Desk and Automation & AI Operations pages — both are quoted in writing after a free scoping call.

How we anonymise these notes