Agent Lightning and Qwen 3.5: coding agent training

The video examines Microsoft's Agent Lightning through a reported Qwen 3.5 9B result on SWE-bench Verified. The speaker cites the V1 paper's increase from 41.8% to 56.4%: a 14.6 percentage point gain, or roughly 35% relative improvement. This does not establish that the model outperforms every frontier coding assistant or will achieve the same results on another repository.

Agent Lightning is presented as open source infrastructure that trains an AI agent from its workflow. Repository edits, tool choices, test results and recovery attempts provide feedback, rather than training only on finished answers. The speaker contrasts this approach with adding prompt rules and choosing a larger model. The described curriculum uses about 6,000 tasks, including roughly 1,000 where Qwen failed all four difficulty probes.

The architecture places a trainer beside the existing agent. An API gateway captures interactions, while a rollout controller launches agents locally or through Kubernetes jobs. This is infrastructure for agent training, rather than a desktop tutorial to run models locally.

The coding example mentions four B200 GPUs alongside a trainer, gateway and model backend. The speaker also warns that weak rewards can encourage agents to chase passing tests at the expense of maintainable code. Stronger models may still suit difficult debugging and architecture work; benchmark gains do not remove human review.