Dec 19, 2025 – Anthropic launches Bloom, an open‑source framework that automatically generates behavioral evaluation suites from a researcher‑specified behavior, speeding up assessments and preventing obsolescence of traditional tests [1].

Dec 19, 2025 – Bloom validates its metrics by distinguishing baseline models from “model organisms” in nine of ten cases and achieving a Spearman correlation of 0.86 with human judgments for Claude Opus 4.1 (versus 0.75 for Claude Sonnet 4.5) [1].

Dec 19, 2025 – The system reproduces and extends prior self‑preferential bias findings, showing that higher reasoning effort reduces bias in Sonnet 4 and that rankings stay stable across configuration changes [1].

Feb 2, 2026 – The AgentRx project releases a benchmark of 115 failed AI‑agent trajectories spanning structured API workflows, incident‑management processes, and open‑ended web/file tasks, each manually annotated with the exact failure step [3].

Feb 2, 2026 – AgentRx introduces a grounded‑theory‑derived taxonomy that categorizes failures across domains, enabling consistent labeling and cross‑domain analysis [3].

Feb 2, 2026 – Using constraint‑violation logs and an LLM‑based judge, AgentRx automatically pinpoints critical failure steps, outperforming prior baselines in both step localization and failure‑type attribution [3].

Feb 15, 2026 – CorpGen defines Multi‑Horizon Task Environments (MHTEs), problem instances that require coherent execution of 45 + interleaved long‑horizon tasks, each lasting 500–1500 + steps, to mirror real‑world organizational work [2].

Feb 15, 2026 – Researchers identify four failure modes (context saturation, memory interference, dependency complexity, reprioritization overhead) that cut baseline autonomous‑agent completion rates from 16.7 % to 8.7 % as task load rises from 25 % to 100 % [2].

Feb 15, 2026 – CorpGen introduces architecture‑agnostic mechanisms—hierarchical planning, sub‑agent isolation, tiered memory, and adaptive summarization—to counter those failure modes and improve task coherence [2].

Feb 15, 2026 – Empirical tests on the OSWorld Office environment show CorpGen delivering up to 3.5× performance gains, achieving 15.2 % task completion versus 4.3 % for baselines, with stable results under increasing load [2].

Feb 15, 2026 – Ablation studies reveal that experiential learning is the primary driver of CorpGen’s advantage; removing it sharply reduces performance gains [2].