Anthropic’s long-horizon agent harness

Following yesterday’s outside look at harness regressions and machine-checked release gates, Anthropic’s Lance Martin explains how the company builds harnesses for long-horizon tasks. The talk focuses on practical choices for builders.

NexForge generates training data for agents

arXiv cs.AI

NexForge generates agent training data and reports results that surpass proprietary models. It puts data generation beside the harness as another system-level source of capability.

A benchmark built around payment integration

arXiv cs.AI

Alipay-PIBench tests coding agents on complex software work such as payment integration. It targets tasks that broad coding scores can miss.

Verbatim chunks beat extracted memory artifacts

arXiv cs.AI

In a controlled comparison, verbatim chunks beat extracted artifacts by 15.9 points on LoCoMo and 22 points on LongMemEval-S. The authors recommend keeping structured artifacts alongside source text instead of using them as a replacement.

Vocabulary dropout targets diversity collapse

arXiv cs.AI

This paper proposes vocabulary dropout to address diversity collapse in advanced model training and curriculum design. The mechanism targets builders working with self-play and agentic training systems.

How software engineers are adapting to AI

The Guardian

The Guardian reports on how software engineers are responding to AI-driven changes in their work. Its focus on collective action adds the workers’ perspective to recent layoff and restructuring coverage.

Robot corrections that persist across sessions

arXiv cs.RO

PhysClaw-0 retains language corrections across sessions while collecting robot-training data. The paper reports gains in data efficiency and task success from that retained guidance.

RouterVLA reuses smoke-test data to select policies

arXiv cs.RO

RouterVLA uses smoke-test data as supervision for choosing vision-language-action policies. The paper reports a 14.64 percentage-point improvement in policy routing across robot deployments.