09/03/2026
𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲: 𝗛𝗼𝘄 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗔𝗿𝗲 𝗥𝗲𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗣𝗮𝗿𝗮𝗱𝗶𝗴𝗺
This new paper by Zhenfeng Cao presents the agent as the software itself and moves the human to intent architect. We read it closely. Five takeaways.
𝟭. 𝗖𝗼𝗱𝗲 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝗮𝗻 𝗶𝗻𝘀𝘁𝗿𝘂𝗺𝗲𝗻𝘁, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺. Traditional software encodes every decision before runtime. Agentic software generates decision logic on demand, writes code for the specific path, and discards it. The capability persists. The code does not.
𝟮. 𝗧𝗵𝗲 𝘀𝗰𝗮𝗹𝗶𝗻𝗴 𝗮𝗿𝗴𝘂𝗺𝗲𝗻𝘁 𝗶𝘀 𝗮𝗯𝗼𝘂𝘁 𝗰𝗲𝗶𝗹𝗶𝗻𝗴𝘀. Human cognitive capacity is fixed. Model capacity grows with compute. The paper frames this as the structural reason agents change which problems are economically tractable.
𝟯. 𝗧𝗵𝗲 𝗵𝘂𝗺𝗮𝗻 𝗿𝗼𝗹𝗲 𝘀𝗵𝗶𝗳𝘁𝘀 𝘁𝗼 𝗶𝗻𝘁𝗲𝗻𝘁, 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲, 𝗮𝗻𝗱 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝗰𝗮𝗹𝗶𝗯𝗿𝗮𝘁𝗶𝗼𝗻. Code production gets commoditized. Defining what good looks like, and building the evaluation loop an agent can self-correct against, becomes the differentiating skill.
𝟰. 𝗧𝗵𝗲 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲 𝗵𝗮𝘀 𝗮 𝗰𝗹𝗶𝗳𝗳. On EvoClaw, agents score above 80% on isolated tasks and at most 38% on continuous evolution across commit histories. The paper names the causes: context drift, error propagation, no model of technical debt, and tests that pass while semantics break.
𝟱. 𝗧𝗵𝗲 𝗴𝗮𝗽 𝗶𝘀 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴, 𝗻𝗼𝘁 𝗽𝗵𝘆𝘀𝗶𝗰𝘀. The author attributes the drop to context management, memory architecture, and verification. Those are buildable.
That last point is the whole reason CDE exists. Information architecture keeps context coherent across sessions. Quality gates catch error propagation before it compounds. Verification is the harness, not an afterthought. The model is the reasoning engine. We build the system around it.
Link in the comments.
