@Xudong07452910: Recently, while looking into agent self-evolution research, I've been focusing on a question: When are the trajectories left by agents worth continued learning for the next round of training? This time, I took a self-evolving agent paper from SEED and ran it entirely in Apodex…

Recently, while looking into agent self-evolution research, I've been focusing on a question: When are the trajectories left by agents worth continued learning for the next round of training? This time, I took a self-evolving agent paper from SEED and ran it entirely in Apodex 1.1. Initially, I had it break down the paper's methods, experiments, and key claims, and map the conclusions to specific figures, tables, and evidence. After the task had been running for a while, I added an additional requirement midway: Focus on checking whether 'incorrect trajectories are summarized into incorrect hindsight skills, which are then continuously reinforced in subsequent training,' while preserving the previously completed research results. I recorded the entire process. What I'm particularly interested in is how agents continue to maintain a long-running task after changes occur. Apodex 1.1 has defined the agent's 'capability units' fairly comprehensively this time: From understanding the goal to actually accessing files, search, and code environments for execution; maintaining task state during the process, receiving new feedback, adjusting based on changes, and ensuring that the delivered conclusions can be re-verified. This is backed by an asynchronous Agent Team. Different subagents can advance in parallel, with intermediate results continuously returned to a Shared Task State, and users can also intervene during task execution. In the context of agent self-evolution, this layer is actually very important. Trajectory learning, reflection, and skill learning all ultimately depend on the experience left behind during execution. If the task state, evidence sources, and failure branches are themselves unstable, training on these trajectories later can easily introduce noise. Apodex extends this line further to AI4AI: using automated processes to connect problem setting, trajectory filtering, training, evaluation, and reflection. In official experiments, they used this process to train Qwen3.5-0.8B, and after 10 rounds of automatic iteration, the evaluation score improved from 51.0% to 56.0%. So, from this experience, I'm more focused on a fundamental question: How do agents leave behind sufficiently reliable, trackable, and verifiable execution processes so that these experiences can truly be reused later? The related models and frameworks have all been open-sourced: FrontierAgent is an open-source agent harness that supports ReAct / Agent Team. Those working on agents can check it out on GitHub: https://github.com/ApodexAI/FrontierAgent… Apodex 1.1 mini (35B) weights have also been released on Hugging Face, and can be deployed locally and paired with FrontierAgent: https://huggingface.co/collections/apodex/apodex-11… The full Apodex 1.1 workbench is now live. The most direct way to experience it is to throw in a paper you're actually reading, a dataset, or some code, and then change the requirements midway through the task to see how it continues to execute: https://apodex.ai