Yuxuan Zhang — BlogNotes on agentic AI, LLM training, and coding agents.2026-08-11T00:00:00+00:00https://yuxuan.world/blog/Yuxuan Zhanghttps://yuxuan.worldLong-Horizon Agents: State, Planning, and Verifiable Hours-Long Execution2026-07-12T00:00:00+00:002026-07-17T00:00:00+00:00https://yuxuan.world/blog/long-horizon-agents/Long-horizon ability is not a longer context window or timeout. This note separates task envelopes, then folds execution state, planning, and recovery into one falsifiable protocol for agents that work sustainably.AutoResearch: From a Falsifiable Research Loop to the Evidence Map2026-07-10T00:00:00+00:002026-07-17T00:00:00+00:00https://yuxuan.world/blog/autoresearch/Autonomous research as an auditable system: a research loop defined by frozen acceptance and clean reruns, then deep research, five research threads, and a project roadmap converged into one falsifiable evidence map.Web-Agent Environments: Turning Browser Use into Verifiable RL2026-07-10T00:00:00+00:002026-07-17T00:00:00+00:00https://yuxuan.world/blog/web-agent-environments/A design note for ClawBench V2 and WebsiteBench: compare self-hosted web, live web, desktop, and mobile environments, then fold tool governance, evaluation integrity, and multi-agent collaboration into one verifiable RL protocol.RSI: A Conditional Theory, Not an Achieved Milestone2026-07-02T00:00:00+00:002026-07-17T00:00:00+00:00https://yuxuan.world/blog/rsi/Recursive self-improvement is not a binary milestone. Starting from Good, the Godel Machine, and modern bounded optimization, this post explains how to state, test, and limit RSI claims rigorously, and folds the agent-skill lifecycle in as the controlled, bounded end of the self-improvement spectrum.Agent Research Environments: What Is Actually Being Measured?2026-07-02T00:00:00+00:002026-07-17T00:00:00+00:00https://yuxuan.world/blog/env/A measurement map for agent environments: what each benchmark and harness actually measures, when a world model counts as an environment, and when synthetic instances are worth training on.Core LLM RL Algorithms: PPO, DPO, and GRPO2026-04-29T00:00:00+00:002026-07-17T00:00:00+00:00https://yuxuan.world/blog/rl/A compact guide to PPO, DPO, GRPO, reward models, critics, online RL, and offline preference optimization for LLMs.Agent Harness Engineering: A Source-Referenced Comparison of Coding Agents2026-04-23T00:00:00+00:002026-08-11T00:00:00+00:00https://yuxuan.world/blog/harness/A bilingual, interactive deep-dive into hermes-agent, claw-code, codex, opencode, openclaw, and pi — with flowcharts tied to source files and selected line references. Now also folds in harness-engineering (six accountability planes: tool contracts, state, authority, control, evidence, evaluator integrity), agentic RL (the harness as part of the policy; PPO / GRPO / SFT+RL on a shared rollout ledger), core LLM RL algorithms (PPO / DPO / GRPO with reward models, critics, and verifiable rewards), and a coding & implementation reference appendix (Classic Hot 100 + from-scratch PyTorch MHA / Conv1D / Conv2D).