XNews

Agent-driven automation of AI research and developer workflows

Agent-driven automation of AI research and developer workflows

Cursor automates AI research with agent fleets. - Outer loop: user feedback and A/B tests fed back - Inner loop: improve RL environments & learning methods - Composer 2.5 became top product model via new training - Agent fleets run experiments, monitor via SSH - Agents notify humans via Slack or PagerDuty - Researchers can launch experiments from Slack - Models distill judge/reward models for evaluation - Recursive improvement accelerates training cycles - Cursor+SpaceXAI trained Grok 4.5 on interaction data - Cursor restricts web access and deletes git history - Cursor maintains private CursorBench task set - Kanban Code turns Claude Code tasks into Kanban cards - Built-in terminal shows task outputs without local switching - Observers report researchers supervise agents more than typing

Key insights

Cursor defines outer and inner training loops

Cursor presented a two-loop framework: an outer loop that collects real user feedback, online metrics and A/B results to feed subsequent training rounds, and an inner loop that improves training efficiency by generating harder RL environments, refining learning methods, creating realistic software-engineering tasks, and building auxiliary models (judge, reward model). Cursor stated that the simplistic formula “more compute → better model” is superficial and emphasized both loops as core to rapid iteration.

Agent fleets automate experiment lifecycle and researcher workload

Cursor described large-scale agent fleets (a primary agent plus many sub-agents) that run continuously to launch, monitor and maintain experiments: agents collect status via SSH, maintain health checks, parallelize tasks and proactively notify humans through Slack or PagerDuty when intervention is needed. Researchers can start and manage experiments from Slack, reducing manual babysitting of training jobs and freeing researchers to focus on higher-level ideas.

Recursive model improvement: models help train better models

Cursor reported that stronger main models can distill improved auxiliary models (judges, reward models), which in turn raise evaluation quality, reward signals and data-generation efficiency. This creates a verified positive feedback loop where models help produce better training conditions for the next generation, accelerating the training cycle without invoking fully autonomous RSI.

Cursor training outcomes and Grok 4.5 collaboration

Cursor stated it has run large-scale model training for about a year and credited Composer 2.5’s product popularity to more RL environments, higher task difficulty and new training methods. Cursor also reported a joint training effort with SpaceXAI on Grok 4.5, using extensive Cursor real-interaction data (code operations, tool usage, developer-agent traces) and reinforcement learning in hard real-world environments; prior models were used to accelerate the next-generation model.

Developer workflow tools and signs of agent-driven labs

A developer tool, Kanban Code, converts Claude Code tasks into Kanban cards that auto-move through stages (todo→in-progress→waiting→review→done), includes an embedded terminal for task outputs, and pushes mobile notifications. Separately, observers reported that top AI labs have researchers who no longer type much and instead supervise concurrent agent tasks, making human typing a new bottleneck in some environments.

Evaluation safeguards against online-answering and dataset leakage

Cursor reported measures to preserve evaluation integrity: restricting network access during evaluation, deleting git history to prevent leakage, and maintaining a private real-code task set (CursorBench) to ensure assessments reflect real capability rather than web-search or dataset leakage.

Source posts