Microsoft has released its agent reinforcement learning framework, Agent Lightning v1.0, on GitHub. The framework centers on a production harness that handles context building, tool execution and the agent-environment interaction loop, while the training engine observes only LLM Q&A logs. Microsoft said it used 6,000 training examples to train a Qwen3.5-9B model and raised its SWE-Bench Verified score to 56.4 from 41.8. Researchers cited benefits and cautioned about reward design and infrastructure needs.

Ankit Jain (안킷 제인), co-founder and CEO of developer productivity platform Aviator, said that as AI writes most software code, engineers will shift from writing code to building the machines that write it. He expects roles to split into product engineers and platform engineers, with growing demand for the latter. Jain said faster code generation is creating friction as teams build separate prompts, guardrails and dashboards without sharing. He said all software companies will effectively become development tools companies.