The clean architecture

Important ideas:

  • Layers
  • Directionality of dependencies
  • Isolate business entities from externalities like frameworks, databases, connected systems, etc
  • How information is passed between layers (simple dictionaries, dtos, etc)

Feels like a tasteful way to organize application logic (ie the code we write) for understanding and maintainability.

Source pdf

OpenAI dev day agentic workflows

  • Youtube video: Here’s the dev day video I watched. Neat idea here about how to structure a computation / workflow using agents.

  • Agentic workflows
    • Agent builder
    • Chatkit
  • Agentkit
  • Getting to the goal is broken down into discrete agents
  • Verification of outputs is talked about and is probably super important as we write less code - we should be able to tell what we have is correct and we should be able to refactor (a recent paper about this came our recommended by Mark Brooker)

OpenAI has a nice workflow that creates a pipeline with stages consisting of individual agents that pass output to the next

Chatkit agentkit builder openai screenshot shows a workflow including input from user through various agents leading to an output back to the user hopefully with the information they requested

Back to links :)

  • Claude Skills are awesome, maybe a bigger deal than MCP: Simple markdown files and supporting scripts to extend what an agent is able to do. Simpler than mcp and burns through fewer tokens building a context it sounds like.
  • How FastAPI path operations work: FastAPI is a python web framework that is very light-weight. Great for tiny to small web applications likely. Looks like a nice way to organize the code in a web application and very easy to get going with.
  • System design walkthrough of a web crawler system: I’ve queued up a few of these to watch over the next couple of weeks. It’s fun to think through problems at a high level this way. General flow of thinking through a system like this:
    • review requirements
    • spec thinking about reads, writes, network b/w needs, disk (different problems will have diff resource requirements, and associated pressures on the system)
    • build a dataflow diagram that follows the data through the system as enters, is processed, and is retrieved by users
    • think about how you’d store data (functional areas) and the actors
    • think about machines (io-bound / cpu-bound, ram, etc)

System design architecture diagram: web crawler for the entire internet

Openai dev day : Internal tools and workflows

  • Seems to rely on data being in good shape with all their knowledge, experience, and insight
  • Make everybody on the team as effective as the best ones
  • Skills (in the agent sense) are created and shared with everyone
    • Find top operator
    • Capture how they work
    • Build into tools
    • Skills are collections of markdown files and scripts / tools that an llm can use
  • Video: Demos of tools for go-to-market team, people & hr, and support

Stuff to look into

  • Check out agentkit
  • Check out agent skills as they relate to codex