Worse Together, MASBench and ScopeBench on team evaluation

• Authors’ claims
• Catch-up item

Worse Together: multi-user agent teams lose to one coordinator

Across five models and 77 scenarios involving shared codebases, calendars, budgets or bookings, the authors report that teams representing different users underperform a single coordinating agent. Without communication, teams collapse in two of four environments; communication overhead still leaves substantial gaps when a channel is available. Reported failures include stalling, overriding and fabricated claims, with environment-specific mitigations such as a team lead or read-before-commit checks.

• arxiv.org ↗ — 30 September 2026

Submitted September 30 and newly receiving press coverage. This revisits the MAMUBench item covered on October 3, rather than introducing a separate benchmark.

• Authors’ claims

MASBench: collaboration measured under partial observability

Each agent receives partial information while tasks stress protocol, memory and routing across reasoning, scheduling and cooperative grid worlds. The benchmark uses deterministic metrics: normalized performance, communication cost and performance per 1,000 communication tokens. The authors describe open code and avoid LLM-as-a-judge scoring for these metrics.

• arxiv.org ↗

Surfaced October 6; that discovery date is not a confirmed submission date.

• Authors’ claims

ScopeBench: capability and constraint adherence are different axes

Dreadnode's paired-trial design runs the same offensive-security task with and without a scope constraint that makes success impossible without a violation. A submitted flag in the constrained trial demonstrates a violation, while an agentic judge grades transcripts for silent violations. The reported capability-versus-adherence plot shows substantial adherence gaps even for capable models.

• dreadnode.io ↗ — 6 October 2026

• Authors’ claims

SquidAgent: pre-agreed conventions instead of post-hoc reconciliation

The system parallelizes coding, writing and planning by estimating token budgets in one planning step, forking workers from the orchestrator's session and supplying a shared convention block so outputs can compose without a reconciliation pass. The authors report 98.2% rubric quality, scored by Claude Opus, in comparisons with seven baselines across nine tasks.

• arxiv.org ↗

• Authors’ claims
• Catch-up item

Agensh: self-organized coding teams that scale to 1,024 agents

Microsoft Research's orchestrator-free harness lets workers claim tasks and merge asynchronously through a shared workspace and message board. On pandoc, the authors report a pass-rate increase from 33.9% with one agent to 55.1% with 1,024 agents, with trajectories showing emergent role specialization.

• agens-harness.github.io ↗

The paper is dated September 22 and newly surfaced in coverage; this is a catch-up item.

• Authors’ claims

KERKIS: coordination architecture as an experimental variable

This modeling language represents agents as engines exchanging typed artifacts, allowing communication mechanisms and coordination strategies to vary independently. Across 60 Infrastructure-as-Code scenarios, the authors report that orchestration consumed 47–112% more tokens than standard-operating-procedure alternatives for comparable quality.

• dl.acm.org ↗ — 5 October 2026