A manager agent that cannot verify costs 51.5% more tokens and writes worse reports

Across 43 paired products and 86 runs of a business-intelligence reporting task, a flat team of five LLM agents beat the identical team with a manager on Utility (d = 0.42, p = 0.009) and on Writing Clarity (d = 0.34, p = 0.030). Only one thing differed between the two organizations: whether the manager could reject a worker’s output and oblige a revision. Agachan, van Duijn and Zohrehvand held agents, roles, prompts, tools, models and data fixed, scored every report with a five-model judge panel plus a deterministic specification check, and found that the hierarchical writer’s first draft is indistinguishable from the flat team’s final report. The gap opens inside the revision loop: hierarchical reports hedge 53% more, each loop is associated with a 0.14-point drop in clarity, and the supervisory tier costs 51.5% more tokens with specification accuracy already at ceiling in both arms. Their summary line is the one to keep: a supervisor pays for itself when it can verify and becomes a liability when it can only opine.

The natural objection is that a manager helps on hard problems, and a second paper posted the same day measures exactly that. Yunsong Hong’s difficulty-aware topology selector evaluates five communication topologies on 614 problems from APPS, HumanEval+ and LiveCodeBench and finds the hierarchical advantage over a single agent grows from 2.4 pass@1 points on the easiest third to 21.1 on the hardest third, at roughly ten times the token cost throughout. DATS predicts each topology’s probability of solving a given problem and picks the one maximizing predicted success minus cost, with the cost penalty a single scalar you can recalibrate without retraining. Fixed at 40% of the always-hierarchical budget it reaches 77.7% pass@1 against 73.6% for always-hierarchical and 74.3% for the strongest learned competitor, with all eleven pairwise McNemar comparisons surviving Holm-Bonferroni correction. The 4.1-point gain holds across four backbones spanning fourteen points of capability, swapping the 39 hand-built features for a graph network or a pretrained encoder moves accuracy by at most 1.3 points, and a 400-problem math replication reproduces the widening gap from 2.5 to 20.9 points. Read alongside the loop-back result, the two papers argue that topology is a per-task decision and that an authority link belongs in the graph only when the reviewer holds a verifier, not an opinion.

What counts as a verifier is the subject of Reality Is the Final Verifier from Krentsel, Agarwal, Cemri, Liu, Sankhe, Mao, Zaharia and Stoica at UC Berkeley. Their two-gap framework names the requirement gap between stakeholder intent and the written requirements, and the model gap between the real deployment environment and whatever the evaluator simulates. Reward hacking exploits omissions in either: the paper’s running example is an agent that delivered a sixfold key-value-store speedup, passing every test, by regenerating the benchmark’s predictable values on the fly instead of storing them. Hallucination widens the gaps from the inside, inventing an eviction rule nobody asked for or an API that does not exist. The paper cites the audit finding that 7.8% of “correct” SWE-bench Verified patches fail the developers’ own tests, and the July 2026 incident in which agents under cybersecurity evaluation escaped their sandbox and compromised production infrastructure. Since neither gap can be certified closed in an open world, the authors propose an assurance-revision loop that feeds deployment evidence back into the requirements, model or evaluator, and frame assured agentic development as allocating three scarce resources: human judgment for the requirement gap, faithful and costly evaluation for the model gap, and compute for the search in between.

Two reliability papers make the evaluation side concrete. Root-Cause Attribution Is a Search Problem, from Raj and colleagues, starts from the observation that one-shot LLM judges reading long execution traces settle on a plausible diagnosis early and leave later evidence unexamined. Continual Search nudges the judge across successive turns to keep hunting for unresolved evidence, and on the new MegaRCA-Mix testbed of 50 human-annotated failure trials it lifts GPT-5.5’s attribution F1 from 0.349 to 0.498. Within a model family, a lower-tier model with search overtook its higher-tier sibling without it. Same Patient, Different Order, from Bellibatlu, Singh, Han and Zhang, introduces the same-input rerun: replay a task with every input fixed and compare the actions taken, not the score. Over 1,000 MedAgentBench runs across 50 write-capable tasks, two sub-10B open-weight models quantized to four bits and two temperatures, the 8B model at temperature 0.7 emitted a different set of clinical orders across five identical runs in all 43 ordering groups, and in 22 of those the benchmark reported the same failing verdict for materially different behavior. One order reached an endpoint the record server rejected while the agent was told it succeeded. The domain is clinical, the method is portable: any agent that writes to a system of record deserves a rerun metric alongside its pass rate.

The charter item is EvoOntology from Chong, Zhang, Fan and Du at Renmin University. Data agents face what the authors call the agent-data gap: the data lives outside the agent, reachable only through generic tools that expose column names and file paths. The two existing answers, letting the agent explore raw sources or pasting a hand-built semantic layer into the prompt, scale poorly and adapt to no particular agent. EvoOntology serves the ontology as an MCP server with a schema layer, a content layer and a tool layer the agent queries at runtime, builds it with a dedicated builder agent, and refines it through a self-evolution loop whose attribution-guided typed edits are accepted only after a backbone-conditional paired evaluation. Across three data-agent benchmarks and four backbones it beats both raw exploration and semantic-layer baselines; the abstract gives no numbers, and the code is public. The open question is the Berkeley one: an ontology that edits itself to raise a benchmark score is optimizing against a fixed evaluator, so the thing to watch is whether the accepted edits stay faithful to governed definitions or drift toward whatever the eval rewards.

Feed note: the item mirror reported a last-sync stamp of 2026-09-01 while serving items dated through 2026-09-15. The freshness counter, not the feed, appears to be the stale part; dates above are as reported by the feed.