← October 31, 2026 briefing · added in the 20:00 KST update
Tsinghua and Zhipu's 'HarnessPro' lets agents rewrite their own harness from execution logs, claims 26.1% lower cost per task (report)
Verified (fact)
Researchers from Tsinghua University, Zhipu AI and others introduced 'HarnessPro.' In this approach, an agent analyzes its own execution logs and then rewrites the harness around it (tools, prompts and execution flow) without ground-truth data. According to reports, the pass rate rose by 2 percentage points and cost per task fell by 26.1%. The summary did not detail which benchmark or model conditions produced these figures.
Needs fact-check (claim)
Researchers' claims: harness self-improvement without ground truth, pass rate +2 percentage points, cost per task -26.1%.
Why it matters (interpretation)
It is part of a 'recursive self-improvement' trend that improves both performance and cost by modifying only the harness without touching model weights. It is worth comparing with recent research finding that more complex harnesses only increase costs.
Confidence medium · official source pending
Sources
More about Zhipu AI
-
ByteDance's 'Doubao Work' opens third-party models including Kimi K3.1 Pro, GLM-5.3 and DeepSeek-V4.1-Flash to team and enterprise customers (report)
According to Pandaily, ByteDance's workplace service Doubao Work now lets team and enterprise plans use models from other companies in...
-
Mistral Large 4 follow-up: an improved release with more RL after preview, pitched for cyber defense, scores 48% on Coding Agent Index (reported)
According to TechCrunch, Mistral AI released the weights of Mistral Large 4, a 1T-parameter model, and pitched cybersecurity as a main...
-
KV cache mixing between requests on Cloudflare Workers AI's GLM-5.3 exposed other customers' prompts and outputs in 155 responses (community report)
According to a GeekNews summary, responses from the GLM-5.3 model served on Cloudflare Workers AI contained other customers' prompts and...
-
Shanghai AI Lab releases full version of 'Atria Dawn', a 744B MoE agent model based on GLM-5.2, with 512K context; weights 'coming soon' (reports)
Shanghai AI Laboratory has released the full version of its agent model 'Atria Dawn', the first since the September preview, with a...
-
Paper Instruments releases Ultramarine 1 Flash, post-trained from GLM-5.3 Flash for $300K; claims it beats Claude Fable 5 on some knowledge-work benchmarks (reports)
Paper Instruments, a five-person research team, released Ultramarine 1 Flash, a model specialized for professional knowledge work, on...