Add AutoKernel example for an LLM optimizing GPU kernels

[harmya]

This runs AutoKernel on Modal. We setup Claude to do a long running optimization job on a Llama Model that identifies and optimizes bottlenecked parts of the model.

Here is how the results would look if you ran this:
result.txt

Type of Change

• New example for the GitHub repo

Monitoring Checklist

• Example is configured for testing in the synthetic monitoring system, or lambda-test: false is provided in the example frontmatter and I have gotten approval from a maintainer
  • Example is tested by executing with modal run, or an alternative cmd is provided in the example frontmatter (e.g. cmd: ["modal", "serve"])
  • Example is tested by running the cmd with no arguments, or the args are provided in the example frontmatter (e.g. args: ["--prompt", "Formula for room temperature superconductor:"]
  • Example does not require third-party dependencies besides fastapi to be installed locally (e.g. does not import requests or torch in the global scope or other code executed locally)

Documentation Site Checklist

Content

• Example is documented with comments throughout, in a Literate Programming style
• All media assets for the example that are rendered in the documentation site page are retrieved from modal-cdn.com

Outside Contributors

You're great! Thanks for your contribution.

---

[devin-ai-integration]

Devin Review found 2 potential issues.

View 5 additional findings in Devin Review.

[devin-ai-integration]

🟡 App name missing required example- prefix per repo convention

The internal/CLAUDE.md:148-152 rules explicitly require app names to use example- prefix with kebab-case (e.g., modal.App("example-vllm-inference")). The current code uses "autokernel" instead of "example-autokernel". While some older examples also lack this prefix, it is a clearly documented convention for new examples.

Suggested change

	app = modal.App("autokernel", image=image)
	app = modal.App("example-autokernel", image=image)

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🟡 Container dependency @anthropic-ai/claude-code is not version-pinned

The internal/CLAUDE.md:59-66 and the PR checklist both require all container image dependencies to be pinned to at least SemVer minor version. npm install -g @anthropic-ai/claude-code installs whatever version is latest at image build time. A breaking release of the Claude Code CLI could silently break the agent loop. This should be pinned, e.g., npm install -g @anthropic-ai/claude-code@1.0.3.

Prompt for agents

In 06_gpu_and_ml/agents/autokernel.py line 74, pin the @anthropic-ai/claude-code npm package to a specific version. Check the latest stable version and change the install command from:
  npm install -g @anthropic-ai/claude-code
to something like:
  npm install -g @anthropic-ai/claude-code@1.0.3

Similarly consider pinning the Node.js version (line 72-73 uses setup_20.x which is OK as a major version pin) and the uv installer (line 71 pulls latest). The uv installer is less critical since it's a build tool, but the Claude Code CLI is the primary dependency that drives the agent loop.

---

Was this helpful? React with 👍 or 👎 to provide feedback.