More Agents Aren't Automatically Better

Yishuo Yuan and colleagues at Shanghai Jiao Tong University and Nanjing University tested what happens when one AI agent hands pieces of a job to several others. On two benchmarks where later steps depend on earlier outputs, three generic multi-agent methods scored below a single agent. On Terminal Bench 2.1, the single agent scored 76.71, against 71.49, 68.27, and 67.47. 1

Where the pieces ran independently, two of the three beat the single agent. Yubin Kim and colleagues at Google and MIT found the same split across 260 configurations, from an 80.8% gain on decomposable work to a 70.0% loss on sequential planning, relative to a single agent. 12

Before an HR or learning team buys or builds an agent team, it should map which steps depend on earlier outputs and check whether one agent does the job as well.

Because the evidence comes from software benchmarks and engineering reports, it sets a test for task design, not a verdict on any HR product.

1. 1. Splitting a job lost to a single agent on tightly coupled tasks and won only on sparse ones, in one preprint. On Terminal Bench 2.1 and NL2Repo Bench, where later steps use earlier outputs, TDAG, DynTaskMAS, and Graph-of-Agents (GoA) scored below the single agent. On sparse AgentIF-OneDay, two of three won. 1
2. 2. Adding agents flips sign with task shape across six benchmarks. Kim and colleagues compared one agent with four multi-agent designs across 260 configurations. Gains ran to 80.8% on decomposable financial reasoning and losses to 70.0% on sequential planning, with an overall mean 0.3% below the single agent. 2
3. 3. The best case for several agents is breadth-first research, by one vendor's measure. Anthropic engineers report that a lead Claude Opus 4 with Claude Sonnet 4 subagents beat single-agent Claude Opus 4 by 90.2% on an internal research eval, while calling tasks with many inter-agent dependencies a poor fit. 3
4. 4. Past a point, more agents cost more and score less, in one ablation. On AgentIF-OneDay, the authors' design scored 71.80 with 4 agents and 185M input tokens, and 64.02 with 6 agents and 352M. Anthropic puts multi-agent systems at about 15 times chat tokens. 13
5. 5. Dividing a people-management process across agents does not remove it from the EU AI Act's high-risk list, on Annex III's text. Annex III, point 4 names uses such as recruitment, promotion, termination, and performance monitoring, not agent counts. That reading is Autonoma synthesis. 6

Brief 024 ("Agents as distributed systems, not chatbots") covered coordination, identity, and control for long-running agents. Brief 017 ("The Agent Can Produce Faster Than You Can Verify") covered review burdens when output outruns checking, and Brief 026 ("Redesign the Work, Don't Bolt AI onto Old Jobs") argued for reshaping jobs around AI. All three assume an agent system exists; this Brief asks whether to split a task across agents at all.

None of the six sources studied an HR or learning process. They test software agents on coding, research, and planning benchmarks. Any application to onboarding, recruiting, learning content, or case handling is Autonoma Intelligence's own reading, and each HR example is an illustration, not a finding.

A split pays only when the steps can run apart

Yuan and colleagues model a task as a dependency graph and cut it at "bridge edges," where one subtask hands off to another. Predicting those edges before the work is done is the hard part: a mispredicted edge can leave an agent waiting on a dependency that never arrives. 1

Terminal Bench 2.1 and NL2Repo Bench have long chains where later steps must use earlier outputs, and on both, all three generic methods fell below the single agent. On sparse AgentIF-OneDay, TDAG scored 59.94 and DynTaskMAS 65.85 against the single agent's 53.57, while GoA, with a more elaborate coordination design, scored 52.67. 1

The paper is a preprint, every run used one base model (DeepSeek-V4-Pro), agents were capped at 4, and the best-scoring method, SAIGE, is the authors' own design, one that adds agents only where the task allows. This Brief rests on the generic baselines against the single agent. 1

Kim's wider test points the same way: every multi-agent variant lost 39% to 70% on planning tasks requiring sequential constraint satisfaction, and tasks where a single agent already topped 45% accuracy saw negative returns from extra agents. The authors credit coordination benefits to matching communication topology to task structure, "not from scaling the number of agents." 2

The best case for splitting still carries a coordination bill

Anthropic's engineers, whose employer sells a multi-agent research product, make the strongest case for splitting. A lead Claude Opus 4 with Claude Sonnet 4 subagents beat single-agent Claude Opus 4 by 90.2% on an internal research evaluation, and the post says such systems "excel especially for breadth-first queries" that pursue independent directions. It also says domains with many dependencies between agents are "not a good fit" today, and that in one browsing evaluation token usage by itself explains 80% of the variance. In Autonoma's own reading, this leaves open how much of the 90.2% comes from the split and how much from added spend. 3

The papers price the same trade. Yuan's team writes that beyond a point "the marginal agent appears to add coordination cost," and Kim finds hybrid systems need 6.2 times the turns of a single agent (44.3 against 7.2), while independent systems amplify trace-level errors 17.2 times against 4.4 times for centralized ones. 12

Walden Yan of Cognition, which sells a coding agent, argues from first principles with no new benchmark data. In his Flappy Bird example, two subagents build parts that do not fit together, and the final agent is left to combine two incompatible interpretations of the job. 4

Dividing people decisions adds failure modes without leaving the high-risk list

Mert Cemri and colleagues at UC Berkeley annotated more than 1600 traces from 7 popular multi-agent frameworks and sorted the failures into 14 modes. Failure rates on open-source systems ran from 41% to 86.7% on different benchmarks, so they are not directly comparable. Many failures came from organizational design and coordination rather than individual agents: better role specifications lifted ChatDev's success rate by 9.4%, and task-objective verification added 15.6% on ProgramDev. 5

Yuan's team adds that decomposition can produce failure modes "harder to audit," including local deadlocks, mispredicted dependencies, and opaque coordination, which complicate debugging, accountability, and oversight. 1

The EU AI Act is the one source here on HR work. Annex III, point 4 lists as high-risk areas AI systems intended for recruitment and selection, for decisions on terms of work, promotion, or termination, for allocating tasks by individual behaviour or personal traits, and for monitoring and evaluating performance and behaviour. 6 It says nothing about agent counts, so in Autonoma's own reading a hiring or performance process split across agents sits where a single-agent version would.

Autonoma forecast: Over the next 12 months, new multi-agent results will again split by dependency structure: gains on independent, breadth-first work, and ties or losses against a single agent on chained work. HR and learning vendors selling agent teams without a dependency map or single-agent comparison will find buyers asking for both. 123

• Replication of Yuan's team's results on a second base model, or with more than 4 agents. 1
• Later tests of whether extra agents still turn negative once a single agent passes 45% accuracy. 2
• Vendors reporting a single-agent baseline and token cost beside multi-agent gains like the 90.2%. 3
• Failure rates for multi-agent frameworks falling below the 41% to 86.7% range Cemri's team measured. 5
• A published test of a multi-agent system on an HR or learning process, against a single-agent baseline.
• Your task map: mark each step by whether it needs the previous step's output, then count the chains.
• Your single-agent trial: run one agent on the same cases a vendor demonstrates, and compare accuracy, cost, and review time.

For CLOs and L&D: Before approving a multi-agent build, map the steps. Modules drafted from one outline may split; chained steps belong with one agent. 1

For CHROs: Ask for the single-agent result on your own cases before paying for a team, and treat hiring and performance uses as the high-stakes case. 6

For HR technology and procurement leads: Ask vendors for the dependency map, a single-agent baseline, token cost, and the objective, output format, and task boundaries each subagent gets. 3

For CIOs: Budget for the extra spend: Anthropic reports about 15 times the tokens of a chat, and Kim's hybrid systems took 6.2 times a single agent's turns. 23

For legal and compliance: Do not assume dividing a people-management process takes it off the Annex III list. Point 4 names uses, not agent counts; that reading is Autonoma's own, so have counsel read the text. 6

Weight: Moderate. The strongest objection is Anthropic's 90.2%: a lead agent with subagents beat a single agent by that margin, and as models improve, coordination costs may shrink until splitting wins almost everywhere. 3

The sources answer in part. That 90.2% came on breadth-first research on an internal eval, the independent case this Brief already allows, and the post calls dependent work a poor fit. On model improvements, Kim finds negative returns from extra agents once a single agent passes 45% accuracy, ranked as the paper's best-supported result, and stronger single agents push more tasks past that line; that second step is Autonoma's own reading. 23

No source tests future models. Cognition dated its critique to 2025, and better dependency prediction could cut the mispredicted-edge errors Yuan's team notes. Still, the burden of proof sits with the split. 14

Picture the vendor demo: a cast of named agents, each with a job title, moving an onboarding case from start to finish. The question is what happens when the second agent's output is wrong, and who sees it.

Yuan's team concludes that collaboration should be introduced selectively by task dependency structure, not by default. A single agent is the cheap control: it gives the baseline, shows where it stalls, and marks the steps worth splitting. 1

Decision rule: if a vendor cannot show which steps depend on which, plus a single-agent run on the same cases, do not buy the team; start with one agent and split off only independent steps.