Agent research is putting workplace delivery, context recovery, and security evidence on the same scorecard

DeepDiscovery targets a common failure in coding agents: finding the obvious file while missing registration code, configuration links, tests, or dependency-injection paths. Its Location-Inference method starts with high-confidence task anchors, then expands over explicit dependencies, implicit configuration links, and module proximity under a budget.

The reported gains are practical. In industrial repositories with 2.67 million lines of code and more than 25,000 files, it improves Full Recall Rate by 2.5 to 7.4 percentage points on medium tasks and 1.6 to 9.2 points on large-subproject tasks. On SWE-bench Verified, the equipped system reaches a 78.6% solve rate, 8.2 points above its baseline.

EnterpriseClawBench turns internal enterprise sessions into reproducible tasks with files, expected deliverables, role labels, rules, and text or visual rubrics. The best Lite score is 0.663 across 32 harness-model combinations, which leaves many artifact-delivery and content-quality failures visible. The benchmark also reports cost, runtime, tool calls, and harness-model pairings, so a base model score cannot hide execution problems.

AFTER studies procedural memory as versioned skill files. Static skills add 2.8 accuracy points on average, and one refinement round adds 3.7 to 6.7 points. Skills trained on diverse multi-model traces reach 73.1% cross-model test accuracy, while narrow skill updates can lose accuracy when moved across roles.

The vibe-coding security study supplies the largest warning signal in the period. The authors collect 10,517 mostly AI-authored applications and audit 200 deployed web apps. Human reviewers validate 1,471 exploitable vulnerabilities after deduplication and exploitability checks. Repeated bug types include broken access control, cryptographic failures, injection, secret exposure, placeholder logic, and unfiltered input.

EVerest adds a different kind of security evidence: a public dataset linking requirements, architecture, documentation, and code for an electric-vehicle charging stack. It contains 84 security requirements and 1,445 fine-grained labels. During construction, the authors found and disclosed a real CWE-1295 plain-text token storage weakness.

The SysML v2 fault-localization paper shows how small code models can improve when domain rules are supplied as training data and inference context. A vehicle-domain knowledge graph encodes physical interface and unit-compatibility rules, then guides synthetic fault generation and repair prompts. On the reported evaluation set, Qwen2.5 Coder 1.5B records 95.7% semantic repair accuracy with fine-tuned full-code output and 91.9% with patch output; its baseline semantic repair accuracy is 0.62%.

WisdomAI’s Text-to-SQL post makes a related production claim for enterprise analytics. Its Adaptive Context Engine builds and updates business context using schemas, logs, dbt, LookML, knowledge bases, feedback, and admin review. On five filtered LiveSQLBench query-only datasets, the reported aggregate accuracy is 20% at baseline, 50% after adding knowledge files, and 85% after context learning.