Advancing Developing Growing Agentic AI Together.

• 🪆 Agents Start Improving Models and the Agents Around Them Thinking Machines’ Inkling took a human-set goal, generated training data and evaluations, ran a fine-tuning job through Tinker and loaded the new weights into its agent session. AIDE² goes one level higher, using an outer agent to rewrite the harness of an inner research agent. Its researchers say the resulting system beat a harness refined by humans for two years on held-out benchmarks. Neither is open-ended recursive self-improvement, but both show model and agent optimization shifting from human-only work into agent-run loops and point toward more specialized AI systems built for specific jobs with agents supplying the labor. TechCrunch

• 🦂 OpenAI Builds a Safety Flywheel With Attacker Agents OpenAI trained GPT-Red through self-play to develop prompt injections that stronger defender models still fail to resist. Successful attacks are fed back into production-model training. GPT-Red succeeded in 84% of held-out scenarios against GPT-5.1, compared with 13% for human red-teamers, and OpenAI says the resulting training helped GPT-5.6 Sol cut failures sixfold. Attacker agents are now helping harden the next generation of models. OpenAI

• 🧑‍✈️GitHub Is Turning Copilot CLI Into a Full Coding Agent Seeking to regain momentum and convert enterprise buyers, the latest Copilot CLI release let agents keep working across sessions, remember their tools and settings, split work into separate branches and hand tasks to other agents. GitHub is also adding guardrails to limit what those agents can access and how far they can delegate. The result is a shift from a coding assistant that answers in the terminal to a more persistent system that can manage larger software tasks over time. GitHub releases

Subscribe to the Daily Agentic AI LinkedIn Newsletter →