Agent Trajectory Distillation

Agent trajectory distillation is the agent-era form of Model Distillation / 模型蒸馏 described in 179: 蒸馏风暴：一场无人公开谈论的技术竞赛. Instead of only asking a teacher model for answers or reasoning chains, a model team can use a stronger model to create tasks, operate inside an environment, generate action traces, and leave behind trajectories that train a student model.

The source defines “environment” concretely: a codebase, terminal, editor, compiler, test runner, unit tests, and other task conditions that let an Agent Harness actually execute work. This overlaps with Synthetic Agent Data, but the distillation frame matters when the student is trained to imitate, compress, or internalize behavior from a stronger teacher model rather than only learning from neutral task feedback.

Key Claims

• Agent distillation can include task design, environment setup, intermediate actions, tool calls, failures, corrections, final answers, and evaluator scores.
• A stronger model can be used as question maker, environment builder, trajectory generator, or evaluator; not every evaluator-only use is typical distillation.
• Trajectory distillation makes AI Verification and Agent Environment Isolation more important because the training signal depends on whether the environment exposes real success or benchmark shortcuts.
• The technique can help weaker models acquire task behavior faster, but it can also transfer teacher-model mistakes, habits, refusal styles, or shallow shortcuts.
• The practical moat is the data pipeline: stable access, realistic questions, filtering, rewriting, correction, data mixing, and checking whether training actually improves the student.

Connections

• Model Distillation / 模型蒸馏 — parent technical category.
• Synthetic Agent Data and Environment-Based Agent Benchmarks — adjacent data and benchmark infrastructure.
• Agent Post-Training, Agent RL, and On-Policy Distillation — post-training methods that can use trajectory data.
• Agent Harness, Agent Environment Isolation, and AI Verification — runtime and validation layers.
• Model Distillation Evidence and AI Model Distillation Governance — evidence and governance concerns when trajectories come from closed teacher models.