New synthetic training framework helps language models master agent-like reasoning tasks

Developing AI agents that can reliably handle multifaceted real-world challenges demands training environments that are both realistic and executable. Current progress in building capable LLM agents faces significant hurdles: existing resources tend to be static, lack diversity, or struggle to support the kind of stateful, tool-rich interactions agents need to learn effectively.

What the research introduces#

A team of researchers has developed InternAgentHarness, a synthetic training platform designed to systematically enhance the agent-like capabilities of large language models. Built on top of the InternBootcamp training framework, the system organizes agent training through a four-layer interface that handles prompt engineering, tool execution, interaction management, and performance scoring.

Central to the approach is an automated converter—called BootcampCLI—that translates diverse agentic tasks into a unified, trainable format. This design allows researchers to move beyond conventional fixed benchmarks: when evaluating agents, failure patterns can be captured and converted into new synthetic training examples, filtered rollouts for reinforcement learning, and subsequent re-testing within the same framework.

Testing and results#

The researchers tested their platform across ten distinct agent tasks spanning text-based and vision-integrated scenarios. Using Qwen3-VL-30B-A3B-Thinking as the base model, they applied both supervised fine-tuning and reinforcement learning with InternAgentHarness. Both approaches produced measurable gains in agent performance compared to the unmodified model.

Why it matters#

InternAgentHarness points toward a more practical path for scaling synthetic agent training. By closing the loop—evaluation feeds into task generation, which drives training, which enables refinement—the system creates a continuous cycle that could accelerate progress toward more robust autonomous LLM agents.