AI Agent News, Monday 26 October 2026

Research · The Decoder · 25 Oct

OpenMobile-2 shows broader app training lifts open smartphone agents

THE DECODER reports OpenMobile-2 trains smartphone agents across more than 110 simulated apps and open-sources its models, data, and training setup. The researchers say app variety raised unseen-app success from 60.7 to 67.5 percent, and the larger model hit 59.6 on SPA-Bench, the best result among open models.

Why it matters. For agent builders, this points to broader app coverage in training as a better lever than more data from fewer apps.

Models DevicesPermalink Read at the-decoder.com

Launch · The Decoder · 25 Oct

Nvidia open-sources Flora for cheaper robot video labeling

THE DECODER reports Nvidia has open-sourced Flora, a video labeling system built from three Qwen3.5 models. Nvidia says Flora covers three annotation tasks and can label robot video for about 72 cents an hour, versus roughly $88.50 for GPT-6 Astra in its estimate.

Why it matters. It could lower the cost of preparing robot video data for training and monitoring with an open alternative to commercial services.

Models Infrastructure CompaniesPermalink Read at the-decoder.com

Money · Fierce Healthcare · 21 Oct

Baba raises $40 million series A for AI Medicare and Medicaid care

Fierce Healthcare reports that Baba raised a $40 million series A led by FirstMark Capital, with CVS Health Ventures and Valtruis participating. The company said it will use the funding for research, more health plan reach and faster value-based arrangements.

Why it matters. It signals more work for AI agents that handle paperwork, billing and insurer calls in Medicare and Medicaid care.

Agents at work CompaniesPermalink Read at fiercehealthcare.com

Also picked on Monday

Launch · InfoQ · 25 Oct

Google releases ML Drift for on-device AI inference across GPUs

InfoQ reports Google released ML Drift, an Apache 2.0 open-source GPU compute engine for on-device AI inference across OpenGL ES, OpenCL, Metal, and WebGPU. Google said ML Drift powers LiteRT and that it will no longer update the older TensorFlow Lite GPU delegate.

Why it matters. For on-device apps, ML Drift offers one GPU route across OpenGL ES, OpenCL, Metal, and WebGPU, with Google steering developers from the older TensorFlow Lite GPU delegate.

Devices InfrastructurePermalink Read at infoq.com
Analysis · The Decoder · 25 Oct

Anthropic says one agent often works better than many sub-agents

THE DECODER reports Anthropic says a single agent that loads context as needed often beats multi-agent setups on quality, cost, and latency. Anthropic says newer models can handle complex plans in one loop and has published a GitHub migration kit.

Why it matters. For teams building agents, Anthropic's advice is to test consolidation carefully because single-agent designs can cut latency and cost, but blind merges can break working flows.

Models Agents at work CompaniesPermalink Read at the-decoder.com
Incident · The Decoder · 24 Oct

Fired OpenAI safety researchers dispute leak claims and lack written details

THE DECODER reports that Tomek Korbak, Mikita Balesni, and Jasmine Wang contest OpenAI's leak claims after being fired. In Wall Street Journal interviews, they said OpenAI never identified in writing what they mishandled and warned the case could deter outside safety auditors.

Why it matters. It may make teams less willing to work with outside safety auditors or speak up about how AI models are monitored.

Safety CompaniesPermalink Read at the-decoder.com
Research · Huggingface · 26 Oct

Paper describes Omni-Woodpecker for omni LLM hallucination correction

Huggingface reports a paper on Omni-Woodpecker, a training-free framework for omni LLMs that reduces cross-modal hallucinations by detecting conflicts across modalities and reweighting hidden representations. The abstract says Qwen2.5-Omni-7B gained 12.9 on CMM and 8.6 on AVHBench, while VideoLLaMA2-AV gained 4.6 and 3.9.

Why it matters. This gives omni agent builders a training-free option to cut cross-modal hallucinations on Qwen2.5-Omni-7B and VideoLLaMA2-AV.

Safety ModelsPermalink Read at huggingface.co
Research · The Decoder · 25 Oct

TasteVal results lead Jack Clark to move self-improving AI forecast to 2027

THE DECODER reports that P-Zero Research's TasteVal benchmark found sharp gains in AI systems' ability to design experiments and read results, with performance roughly doubling every three months since December 2025. It says Jack Clark now sees recursive self-improvement as possible by 2027, earlier than his previous 2028 view.

Why it matters. For AI agent builders, the benchmark points to faster progress in models that can plan and assess research work with less compute and cost.

Models CompaniesPermalink Read at the-decoder.com