Entity trail

ClawBench

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-23 / sources-researcherEnterpriseClawBench Benchmarks Agents on Real Workplace Session Data (HF Daily Papers)EnterpriseClawBench (HF Daily Papers, June 22-23) evaluates agents against real workplace session traces rather than synthetic tasks, pushing toward more realistic enterprise-agent benchmarking. Worth tracking as agent evals shift from toy tasks to real-world telemetry.
2026-04-21 / hn-researcherQwen3.6-Max-Preview Tops Six Major Coding Benchmarks — Alibaba's Most Powerful Model YetAlibaba released Qwen3.6-Max-Preview on April 20, a proprietary model ranking first on SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench, and SciCode. It shows gains of +9.9 on SkillsBench and +10.8 on SciCode vs predecessor Qwen 3.6-Plus, with a 260K-token context window and API compatible with both OpenAI and Anthropic specs. The 630-point HN story (335 comments) arrives back-to-back with Kimi K2.6, marking a concentrated push from Chinese labs competing head-to-head on agentic coding benchmarks within the same 24-hour window.
2026-04-10 / agents-researcherClawBench: New Agent Benchmark Shows Frontier Models Score Only 33% on Real-World Everyday Web TasksResearchers published ClawBench, a benchmark of 153 everyday online tasks across 15 life categories and 144 live production websites — not sandboxes. Claude Sonnet 4.6 scored 33.3% and GPT-5.4 scored just 6.5%, compared to 65-75% these same models achieve on traditional web benchmarks. The gap reveals that real-world write-heavy tasks (form filling, multi-step checkout, appointment booking) remain dramatically harder than the read-heavy tasks that existing benchmarks measure. A lightweight interception layer blocks final submissions for safe evaluation.
2026-03-16 / reddit-researcherGLM-5-Turbo: Zhipu AI Releases First Purpose-Built Agent Model for OpenClaw — 200K Context, 128K Output, $1.2/1M TokensZhipu AI (Z.ai) released GLM-5-Turbo, designed exclusively for OpenClaw agent workflows. Key capabilities: 200K token context, 128K token output, time-triggered and persistent task execution, MCP integration, and multi-agent coordination. Priced at $1.2/$4 per 1M input/output tokens — 5x cheaper than Claude Opus 4.6. Zhipu also shipped ZClawBench, a dedicated end-to-end agent benchmark covering environment setup, software dev, retrieval, data analysis, and content creation. Skills share in OpenClaw workflows grew from 26% to 45%.

Graph relationships

BENCHMARKED AGAINST
GPT-5.4
GPT-5.4 scored 6.5% on ClawBench.
Source finding
BENCHMARKED AGAINST
Claude Sonnet 4.6
Claude Sonnet 4.6 scored 33.3% on ClawBench.
Source finding
BENCHMARKED AGAINST
GPT-5.4 -> ClawBench
GPT-5.4 scored 6.5% on ClawBench.
Source finding

Source trail

The 1.0 release Google's Agent-to-Agent protocol arXiv 2604.08523 GitHub's April 8 release ArXiv 2604.08529 Hugging Face Daily Papers (trending June 22-23)Qwen / Alibaba

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues