The frontier is learning to act — Aug 23

Three Aug. 21–22 developments point in one direction: frontier agents are getting richer inputs and more room to act, while the public record for controlling them remains thin.

DeepSeek adds vision to its agent stack. DeepSeek launched the experimental V4-Flash-Vision-Exp on its API platform with mixed text-and-image input, Files API support, and the same text capabilities as V4-Flash, according to DeepSeek. The company says visual-agent performance moves close to Opus-4.8; its update lists Terminal Bench 2.1 at 83.9, DeepSWE at 59.3, and DSBench-Hard at 63.6.12

Faraday trains for research judgment. London lab Inherent says its Faraday agent independently reproduced findings from published scientific papers without seeing the answer in advance. The agent runs on a 27-billion-parameter Qwen 3.6 model and was compared with Claude Opus 4.8 and GPT-5.5; Inherent also trained it with reinforcement learning to choose experiments worth running and design them well.3

Containment plans remain hard to see. Guidelight's review of public materials from Anthropic, Google, OpenAI, Meta, and xAI scored OpenAI highest at 3 out of 5; Anthropic and Meta scored lowest. The assessment measures public disclosure, so the result describes what outsiders can inspect rather than every safeguard inside each lab. OpenAI told TechCrunch that it can restrict permissions, pause workloads, limit deployment, or take a model offline.4