A Kuaishou algorithm expert shared at AICon 2026 how a commercial conversational agent was deployed in e-commerce, balancing effectiveness, latency and cost. The talk breaks down dynamic knowledge base construction, user-simulator training, agent self-evolution and reward assignment under sparse conversational signals.
Self-Evolution
自进化
An AI paradigm where systems continuously optimize themselves via environmental interaction and feedback.
Curated Intelligence
A new paper titled 'How Do Coding-Agent Harnesses Self-Evolve?' reveals that self-evolution mechanisms for AI coding agents, such as post-task retrospectives modifying Skill files, yield limited success. Tested on SWE-bench with Claude Code, only a few of 125 suggestions fully resolved target issues, while some caused redundant checks in unrelated tasks. The study suggests that without clear acceptance criteria, relying solely on AI self-iteration is unstable and requires human guidance.
ModelBest has open-sourced 'ForgeKernel', an autonomous agent that generated a FlashMLA Prefill operator outperforming DeepSeek's hand-written version by 9.8% on NVIDIA B200 after 22.5 hours of self-evolution without human code. This milestone demonstrates AI's capability to autonomously optimize low-level infrastructure, signaling a paradigm shift from manual coding to AI-generated solutions with human architectural oversight.
ModelBest has open-sourced 'ForgeKernel', an autonomous agent for kernel generation. Without access to source code, the system evolved over 22.5 hours to produce a FlashMLA Prefill kernel on NVIDIA B200 that outperforms DeepSeek's human-written implementation by 9.8%, marking a significant milestone in AI-driven infrastructure optimization.
Huawei and partners have released and open-sourced the enterprise-grade AgentOS 'openJiuwen', addressing challenges in scaling agents from demo to production such as complex collaboration, experience retention, and compute efficiency. The platform natively supports multi-agent swarm coordination and workflow orchestration, introduces a self-evolution mechanism based on execution trajectories to reuse business insights, and connects models with compute resources via a plugin architecture while providing enterprise-level security and isolation.
AliExpress Tech published a comprehensive overview of Agent Self-Evolution, addressing the limitations of static foundation models. The article proposes a closed-loop mechanism where Agents dynamically adapt and restructure themselves across multiple dimensions, including Prompts, Skills, memory, and model weights.