Results for “agents”

…Anthropic also published an AG-UI adapter for Claude Managed Agents, mapping chat threads to managed sessions and streaming text, tool calls, and thinking into custom UIs. Model Economics, Usage Limits, and the Enterprise Shift Toward Open …

…was the writeup on stateless decision memory for enterprise agents, which replaces mutable per-agent state with immutable decision logs/event sourcing to improve horizontal scalability, auditability, and fault tolerance @omarsar0. In a…

…that loosely resembles CRISPR. Per @iScienceLuvr, about 950 agents ran for 21 hours and used 210M tokens before one agent flagged the pattern. Humans then carried out Claude-proposed experiments: expression in E. coli plus RNA-seq, which…

…its audit of the summer incident. In that incident, OpenAI agents “escaped containment” and hacked Hugging Face. Monitorability concerns: Korbak says he had spent months raising concerns that labs are losing the ability to monitor agent…

…Agent ReliabilityK3’s early evals are strong, especially for agents/coding: On Agent Arena, Kimi K3 Max reportedly ranks #1 among open-weight models with +9.75% net improvement, leading across multiple signals including confirmed success…

…with Fable delegating to smaller models in Claude Managed Agents @ClaudeDevs Multiple testers described Fable as slow, token-hungry, expensive, but unusually capable: Dan Shipper said it routinely used 500k to 1M tokens on tasks and was…

…32K context, built for Bing retrieval and web-grounding for AI agents (JordiRib1, mustafasuleyman). Google/Gemma updates: googlegemma linked an article; danielhanchen said Gemma-4 finetuning works across 2B, 4B, 26B, 31B in Unsloth;…

…17.7% to 31.4% on CritPt via decomposition into specialized agents. On coding/program synthesis, ProgramBench’s first task was reportedly solved by GPT-5.5 high/xhigh, with xhigh outperforming Opus 4.7 xhigh across metrics. Retrieval and…

…decoding, and prefill/decode disaggregation. Its lead metric is Agents per Megawatt, with early DeepSeek V4 Pro results favoring GB300 and B300 over Hopper and AMD in the tested configs. This is one of the more consequential infra…

…“tasks”, memory, and evals (and the backlash against complexity)Agents “just work” more often now, but still fail off-distribution: Several practitioners describe a step change in reliability/utility for coding agents across recent frontier…

…“basically just LangGraph,” while @huntlovell argued that real agents are fundamentally state machines. The operational side showed up in launches like LangSmith Sandboxes, Agno Environments, and LangChain’s own writeups on IssueBench for…

…multimodal uses that depend on fine visual detail: computer-use agents reading dense screenshots, data extractions from complex diagrams, and work that needs pixel-perfect references. More details in the focused topic summary below. AI News…