Agentic AI is moving from raw model abilities toward shared skills, evaluation frameworks, guardrails and agent-improvement loops, with coding and enterprise workflows as primary testbeds.
The donation of Agent Skills to the Agentic AI Foundation shifts governance of the open standard for agent capabilities and workflows. With more than 70 integrated platforms, the move has direct implications for on-premise deployments: portability, c...
Microsoft Research has introduced Flint-interactive, a language for building interactive data visualizations on top of Flint. It moves interaction logic away from low-level rendering code and toward semantic events plus update operators handled by a ...
The discovery of PHP2Shell with AI shows a structural shift: LLMs turn bug hunting into mass scanning, shrinking the coordinated disclosure window to zero. Chainguard and Amazon found the same flaw in parallel. The bottleneck is not finding bugs but ...
The Mesa project has updated its AGENTS.md file to allow the use of LLMs in drafting code comments and commit/patch messages, provided a human reviews and edits them before integration. A move that signals a governance shift: no longer absolute bans,...
A user tested LFM2.5-24-A2B locally. With llamaampere it hits 130 tokens/s on an RTX A4500 but loops and fails simple requests. In LM Studio chat stabilises, but on the Hermes task for a Garnet plugin it produces broken code, omits functionality, and...
SkillEval introduces an evaluation framework based on skill profiles automatically discovered by LLMs and estimated with cognitive diagnosis models. Tested on 3,811 models and 9,523 items, it produces more calibrated and predictive profiles. Applicat...
Recursive Flint is a coding model with 40 billion parameters and 2.5 billion active. It beats Qwen3.6-35B-A3B on TerminalBench 2.1 and SWE-bench Pro v2, offers a 262k token context and a 3:1 mixture of sliding window and full attention. The architect...
XHToken has published Spark-X2.6-4B, an evolution of Spark-X2.5-4B. Same architecture, gains from improved post-training in coding, agents, long context and multilingual tasks. Three adjustable reasoning levels balance quality, latency and inference ...
Ornith-1.5-122B extends the self-improvement loop of Ornith-1.0: instead of optimizing only scaffold and rollout, it jointly generates tasks, scaffolds, and solutions through reinforcement learning. GGUF weights lower the barrier for local use, but t...
Researchers propose Structured Authority Injection, a protocol that routes human corrections through typed, schema-validated MCP tool calls and maintains a ranked authority context prepended to each LLM prompt. It defines three override classes with ...
Google has open-sourced Harness2, a system that lets an AI agent refine its own harness for each task and retain lessons learned, without touching model weights. Tested with Gemini 3.5 Flash and GLM-5 on legal, office, and professional benchmarks, it...
