Large language model (LLM) agents increasingly rely on reusable skills, but skills are usually written by hand, and letting an agent change them on its own calls for explicit user control. We report on EvoAgent, an evolvable agent framework built for a foreign trade assistant, and share the engineering choices behind it. We organize the agent as a harness of four components (reasoning, context engineering, tools and state) and confine self-improvement to the user's persistent state: profile, memory and skills. Skills are multi-file packages with lazily loaded references, matched in three stages (keyword, embedding, LLM). The main agent can delegate to sub-agents with their own context, and long histories are compressed while an index of recoverable resources is kept. An offline loop reviews completed sessions to update the profile and memory and to propose skills; in the open-source release a skill changes only after the user confirms the proposal. On 172 foreign trade test cases scored by an LLM judge on five dimensions, GPT-5.2 with EvoAgent scored about 28% higher than GPT-5.2 called directly with only the final query. This is a single run judged by the model under test, so it measures the harness as a whole rather than any one mechanism, and swapping in GPT-4.1 or Qwen3.5-35B-A3B lowered the scores relative to calling those models directly. Several mechanisms described here were added after the version we evaluated and are not evaluated. The evaluation data is proprietary, so the results are for reference only. An open-source release of the code, which is not the evaluated business version, is available at https://github.com/Focus-AI-Center/Mentarc-EvoAgent
Decomposition-based Programming-by-Example splits a task into subtasks: a decomposer predicts intermediate subgoals, and a learned synthesizer generates a program for each subgoal. Execution-decomposition approaches train the decomposer to imitate ground-truth (GT) subgoals and thereby treat decomposition quality as a property of the task alone. A learned synthesizer, however, has its own inductive biases, and GT subgoals reflect how the annotator factorized the task, not what this synthesizer can solve. A decomposer that imitates them can therefore propose subgoals that are logically valid but intractable for the synthesizer. Motivated by this observation, we propose Solver-Aware Decompositions (SAD), which keeps supervised training on GT subgoals as a structural scaffold and additionally trains the decomposer online with self-critical policy gradients against a frozen synthesizer. Each sampled subgoal is rewarded with the negative cross-entropy of the synthesizer on the target program, a continuous signal of how difficult the subtask is for this synthesizer. Our experiments show an accuracy paradox: with the synthesizer fixed, a GT-supervised decomposer agrees more often with the GT subgoals, yet its subgoals lead to lower synthesis accuracy, although the synthesizer was trained on the same GT data. On two list-processing domains, SAD improves synthesis and task accuracy over the GT-supervised decomposer, both in-distribution and under length generalization; the same holds under zero-shot transfer to an external benchmark and for a fine-tuned language model. Even an oracle that receives the GT subgoals at inference time, with the same synthesizer and search, fails on tasks that SAD solves, and fewer than 2% of SAD's solutions to these tasks match the GT program. Decomposition quality is therefore not a property of the task: it depends on the synthesizer that executes the subgoals.