Brief №028 Audit Packet
More Agents Aren't Automatically Better
This audit packet supports Brief №028: More Agents Aren't Automatically Better. Read the brief first for the full argument.
Autonoma briefs are designed to be inspectable. This packet shows what the brief claims, how each claim was tested, what it does not claim, and where caveats remain — without exposing raw internal logs, prompts, source-routing mechanics, hashes, local paths, secrets, or unpublished candidate claims.
Brief Summary and Audit Verdict
Brief 028 asks whether an HR or learning team buying or building an agent team has mapped which steps depend on which, and whether one agent does the job as well. It answers that splitting work helps when parts run independently and hurts when each step depends on the last.
Audit verdict: Supported at the benchmark and engineering-report layer. Yuan's preprint holds the head-to-head scores, Kim's paper the 260-configuration spread, Cemri's paper the failure taxonomy, and Anthropic's post the 90.2% result. The Brief withholds any claim that these software results transfer directly to HR and learning processes, as no source studied one. The HR application, Annex III reading, and 12-month timing are Autonoma Intelligence synthesis.
Claim Register
| # | Claim in the Brief | Status | Domain | Bound |
|---|---|---|---|---|
| B028-C01 | On Terminal Bench 2.1 and NL2Repo Bench, TDAG, DynTaskMAS, and GoA scored below a single agent (76.71 against 71.49, 68.27, and 67.47 on Terminal Bench 2.1). On AgentIF-OneDay, TDAG and DynTaskMAS beat it (59.94 and 65.85 against 53.57); GoA did not (52.67). [1] | Supported | arxiv.org | Preprint; one base model, at most 4 agents. |
| B028-C02 | Across 260 configurations, change against a single agent ran from an 80.8% gain on decomposable financial reasoning to a 70.0% loss on sequential planning, with a mean 0.3% below it. Tasks where one agent topped 45% accuracy saw negative returns from extra agents. [2] | Supported | arxiv.org | Preprint; some benchmarks use 20-instance subsets. |
| B028-C03 | The authors' design scored 71.80 with 4 agents and 185M input tokens and 64.02 with 6 agents and 352M. Hybrid systems need 6.2 times a single agent's turns, and Anthropic puts multi-agent systems at about 15 times chat tokens. [1][2][3] | Supported | arxiv.org; www.anthropic.com | One ablation; the 15 times figure is the vendor's. |
| B028-C04 | A lead Claude Opus 4 with Claude Sonnet 4 subagents beat single-agent Claude Opus 4 by 90.2% on Anthropic's internal research eval, strongest on breadth-first queries. The post calls heavy inter-agent dependencies "not a good fit" and says token usage explains 80% of variance. [3] | Supported | www.anthropic.com | Vendor's own eval; breadth-first research only. The 80% figure comes from BrowseComp. |
| B028-C05 | Cemri and colleagues at UC Berkeley annotated more than 1600 traces from 7 popular multi-agent frameworks into 14 failure modes, with failure rates from 41% to 86.7%; better role specifications lifted ChatDev's success rate by 9.4%. [5] | Supported | arxiv.org | Open-source frameworks; rates not directly comparable. |
| B028-C06 | Yan's Flappy Bird example has two subagents build parts that do not fit. [4] Yuan's team says decomposition can produce failure modes "harder to audit." [1] | Supported | cognition.ai; arxiv.org | [4] is an essay with no data; [1] is an impact statement. |
| B028-C07 | Annex III, point 4 lists as high-risk areas AI systems intended for recruitment and selection, and for decisions on terms of work, promotion, or termination, for allocating tasks based on individual behaviour or personal traits, and for monitoring and evaluating performance and behaviour. [6] | Supported | eur-lex.europa.eu | Names uses, not agent counts; no dates, obligations, or fines stated. |
| B028-C08 | Point 4 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. [6] | Autonoma synthesis | - | A reading of the text. |
| B028-C09 | Adding agents improves results by default. The Brief rejects this: collaboration should be introduced selectively, based on the dependency structure of the task. [1][2] | Rejected by the Brief | arxiv.org | Benchmark work only. |
| B028-C10 | Any Deep Research Bench II single-agent figure from [1]. | Withheld | arxiv.org | Not used. The paper reports two different values for it. |
| B028-C11 | The benchmark results transfer directly to HR and learning processes. | Withheld | - | No source studied an HR or learning process; HR examples are illustrations. |
| B028-C12 | Over the next 12 months, multi-agent results will again split by dependency structure, and HR and learning vendors selling agent teams without a dependency map or single-agent comparison will find buyers asking for both. [1][2][3] | Autonoma synthesis | - | Labeled forecast. |
Source Ledger
- [1] Yishuo Yuan and others, arXiv preprint 2609.19759, 17 September 2026. Holds the benchmark scores, ablation, and failure-mode warning. Proves nothing about HR work, other base models, or over 4 agents in the main comparisons. https://arxiv.org/abs/2609.19759
- [2] Yubin Kim and others, arXiv preprint 2512.08296, 9 December 2025. Holds the 260-configuration spread, turn and error multipliers, and the 45% result. Does not reach beyond its six benchmarks. https://arxiv.org/abs/2512.08296
- [3] Jeremy Hadfield and others, Anthropic Engineering, 13 June 2025. Holds the 90.2% gain, its breadth-first scope, the 80% variance figure, and the 15 times token figure. Cannot separate the split from the spend. https://www.anthropic.com/engineering/multi-agent-research-system
- [4] Walden Yan, Cognition blog, 12 June 2025. Holds a practitioner argument and the Flappy Bird example. Offers no benchmark data. https://cognition.ai/blog/dont-build-multi-agents
- [5] Mert Cemri and others, arXiv preprint 2503.13657, 17 March 2025. Holds the 14-mode failure taxonomy, the 41% to 86.7% range, and the ChatDev and ProgramDev gains. Open-source frameworks only. https://arxiv.org/abs/2503.13657
- [6] European Parliament and Council, Regulation (EU) 2024/1689, Annex III, Official Journal, 12 July 2024. Holds the point 4 wording. Silent on agent counts. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Two of the six sources are vendor posts: [3] from Anthropic, which sells a multi-agent research product, and [4] from Cognition, which sells a coding agent.
Evidence Boundaries
- Every source covers software agents; none studied an HR or learning process, so the HR applications are Autonoma Intelligence's own reading.
- [1] is a preprint on one base model (DeepSeek-V4-Pro) with at most 4 agents; its best method, SAIGE, is the authors' own, so the Brief rests on the generic baselines.
- [6] supplies point 4 only; the step that dividing a process across agents does not leave the high-risk list is Autonoma synthesis, and the Brief states no dates, fines, or obligations.
Dissent and Limiting Case
Weight: Moderate. The strongest objection is Anthropic's 90.2% result and the argument that better models will erase coordination costs. [3]
That result is breadth-first research on the vendor's internal eval, and the post calls dependent work a poor fit. Kim finds negative returns once a single agent passes 45% accuracy, so stronger single agents leave less room for a split; that second step is Autonoma's own reading. No source tests future models. The claim survives narrowed: the burden of proof sits with the split. [1][2][3]
Falsification
This Brief is wrong, or has to be rewritten, if:
- Yuan's setup, repeated on other base models or with more than 4 agents, shows generic multi-agent methods beating a single agent on tightly coupled tasks;
- later tests find extra agents still help once a single agent passes 45% accuracy; or
- a published test on an HR or learning process shows a split beating a single agent on chained steps.
Forecast Label
The Brief's forecast reads: "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." It is Autonoma Intelligence synthesis. The timing is not a fact quoted from any source above.
Methodology
This audit checks each figure in Brief 028 against the source that states it, with the EU AI Act text read in its Official Journal version. The HR application, the Annex III reading, and the 12-month timing are Autonoma Intelligence synthesis. This draft was prepared with an AI assistant and reviewed and approved by the editor.