AI agents improve when training tasks adapt to their weaknesses

The authors report gains across three tool-use benchmarks by repeatedly tailoring tasks and teacher demonstrations to an agent’s changing capabilities.

Chinese Tech

Xuanang Gao · Kepeng Lin · Xinyan Shi · Yong Deng · Shuheng Shen · Changhua Meng · +3 more

Shanghai Jiao Tong University · Ant Group · Huazhong University of Science and Technology · Nanjing University of Aeronautics and Astronautics

Research Digest·2 November 2026·2 min read

Illustration: AI agents improve when training tasks adapt to their weaknesses

In one line

EvoMCP improves AI agents by co-evolving training tasks and demonstrations based on student feedback.

Before you read this

Three ideas this paper builds on

self-evolving agentAn agent that improves by reusing its own experience, memory, or self-generated feedback, with success depending on verification and design.81 papers →skill evolutionThe process of developing, refining, or optimizing reusable capabilities in LLM agents through learning, adaptation, or experience accumulation.40 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.59 papers →

Read it as

AI agents that use tools often learn from a fixed set of examples, which may not address their weaknesses as they improve. This means their training data stays the same even as their skills change.

EvoMCP solves this by creating a loop: the agent's performance is checked, and new training tasks are designed to target its specific gaps. It also improves example solutions, fixing failed attempts and polishing successful ones.

Tested on three benchmarks, EvoMCP raised scores by 16.5%, 10.2%, and 12.6% over the starting agent. This shows that training data that adapts to the learner's needs can make AI agents much more effective at using tools.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

self-evolving agent supervised fine-tuning curriculum learning skill evolution synthetic data generation