Research·261 stories·Updated October 5, 2026
More from Research
Research
Harness Design Reshapes Coding Agent Performance Across Models
A September 2026 empirical study published on arXiv found that harness design choices—including planning strategies, action-space constraints, and context management—significantly altered how coding a
Research
PETSCAgent-Bench: Frontier AI Models Struggle With Scientific Code
Researchers at Argonne National Laboratory and collaborators released PETSCAgent-Bench on September 10, 2026, a new evaluation framework showing that state-of-the-art language models can generate read
Research
Scientific agents cluster at same score level across blind research ta
A September 2026 benchmark paper found four coding agents scored within a narrow 58.4–60.3/100 range on 40 blind scientific discovery tasks, with no statistically reliable separation and critical gaps
Research
Anthropic's AI agents outperform 28 human alignment researchers
Anthropic reported on August 28, 2026, that automated alignment researchers built on Claude beat proposals from 28 experienced safety experts across all seven evaluated alignment failures, improving s
Research
DAREBench: New 233-Task Benchmark Reveals Agent Model Trade-offs
Researchers introduced DAREBench on September 5, 2026, a deployment-aware evaluation framework that tested 35 AI models across 233 agent tasks and found no single model dominates all workload groups.
Research
Agents Are Systems, Not Models—New Benchmark Shows Setup Beats Model C
Researchers released a benchmark across four scientific tasks on October 1, 2026, finding that roughly 54% of agent outcome variance comes from system configuration rather than the underlying model. T
Business
Enterprise AI agent adoption hits 62% in Q3 2026
KPMG's Q3 2026 Global AI Pulse survey found that 62% of organizations across 20 countries are now building, deploying, or developing AI agents, with multi-agent systems jumping to 25% adoption—the sha
Launches
Google rolls out Gemini 4 Argon to cybersecurity partners
Google announced Gemini 4, code-named Argon, on September 30, 2026, beginning a limited rollout to select cybersecurity partners after months of delays. The flagship model is positioned to compete wit
Research
Sierra open-sources Hyper-τ-Bench for agent construction
Sierra released Hyper-τ-Bench on September 8, 2026, a benchmark that evaluates whether AI coding agents can not only act as agents but also construct them. The open-source evaluation places a develope
Research
MIT, Sakana AI cut coding-agent eval costs with SIFT framework
Researchers at MIT and Sakana AI have released Self-Improvement via Fast Tree-Search (SIFT), a framework that uses an LLM judge to dramatically reduce evaluation costs for self-improving coding agents
Research
Misaligned agents need coalitional safeguards, study finds
Researchers at the University of Pennsylvania and University of Maryland published a framework on September 14, 2026, showing that delegating authorization decisions to potentially misaligned AI agent
Research
Agent Leaderboard Rankings Unreliable Without Task Diversity
Researchers on September 30 published a Bayesian framework showing that adding more evaluation tasks does not automatically stabilize agent rankings, undermining confidence in current leaderboard comp
Research
CompMat-Bench: 94-Task Benchmark Tests AI Agents on Materials Science
Researchers introduced CompMat-Bench on September 30, 2026, a benchmark of 94 computational materials science tasks designed to evaluate AI agents without running expensive simulations. The benchmark
Research
Cisco Survey: Over Half of Enterprises Running Agentic AI in NetOps
Cisco and Omdia released joint research on September 23, 2026, showing that more than half of 1,000 surveyed IT leaders have already deployed agentic AI systems in production network operations, while
Research
Safety vs. Task Completion Trade-off Found in LLM Agents
Researchers at seven institutions have published a new framework showing that tool-using LLM agents often achieve high safety rates at the cost of failing to complete authorized tasks. The study, rele
Research
Randomized Audits Can Backfire Against Deceptive AI Agents—Study
A new arXiv paper posted September 29, 2026, shows that stronger auditing of AI agents capable of concealing misconduct can paradoxically make violations harder to detect unless specific safeguards ar