@vintcessun: I always thought AI agents could only write ordinary code. Turns out MIT HAN Lab is directly using an agent workflow to design and optimize CUDA kernels. Hand-tuning is time-consuming and easy to miss solutions. They came up with a workflow of "task contract + agent loop + small-step verification", letting the agent research, implement, verify...
Summary
MIT HAN Lab proposes a method to automatically design and optimize CUDA kernels using an AI agent workflow. Through a process of task contracts, agent loops, and small-step verification, the agent can autonomously iterate and optimize within a specialized toolchain, replacing manual tuning.
Similar Articles
@AxtonLiu: https://x.com/AxtonLiu/status/2073791557547794579
This article discusses the concept of Agent OS, emphasizing the division of tasks into multiple workstations (fetch, refine, verify, confirm) through specialization, each managed by an independent Agent to achieve controllable automation. The author uses the example of digesting browser tabs to demonstrate how specialization isolates context, responsibility, and risks, ensuring the accuracy and reliability of AI output.
This article systematically reviews AI Agent architecture and engineering practices, covering control flow, context engineering, tool design, memory, multi-agent organization, evaluation, tracing, and security. It is based on the OpenClaw implementation and emphasizes the critical role of Harness (testing and validation infrastructure) for system stability.
This article systematically reviews AI Agent architecture and engineering practices, covering control flow, context engineering, tool design, memory, multi-agent organization, evaluation, tracing, and security. It is based on the OpenClaw implementation and emphasizes the critical role of Harness (testing and validation infrastructure) for system stability.
@jakevin7: Sharing something interesting Maka is currently working on: letting agents automatically optimize their own system prompt, fully closed-loop, without any human intervention. Karpathy's autoresearch, AEGIS, etc. have explored similar directions—a goal-driven self-reinforcement learning system.
Maka is a local-first desktop AI workbench whose new feature allows agents to automatically optimize their own system prompts by generating variants, using Harbor container evaluation, and an acceptance policy for iterative improvement, all without human intervention.
@jinchenma_ai: https://x.com/jinchenma_ai/status/2061835131107860582
The article proposes an engineering methodology based on AI Agent (Skill), suggesting that deterministic tasks be solidified into scripts to reduce new decisions made by the large model at runtime, thereby improving stability and token efficiency. Taking video subtitle processing as an example, it demonstrates a four-step engineering process.
@vintcessun: Tonight I came across a learning roadmap project that redefined where to start learning Agent. I used to think Agent was just a pile of tools and frameworks, but its core is the "observe-think-execute" loop and the harness engineering's organization of permissions, state, and backtracking. It breaks down learning into building a minimal Agent loop from scratch all the way to deploying a real Agent, with 8 stages, each with clear deliverables and recommended resources — not just links but an actionable todo list. This systematic approach made me realize my previous learning was too fragmented.
An open-source learning roadmap project called Agent-Learning-Hub, which breaks down AI Agent learning into 8 stages from building a minimal Agent loop to production deployment, providing executable todo lists and recommended resources, maintained by members of the Datawhale community.