• A skill to create new skills: Interesting to see what teaching an agent to use a tool or follow a process looks like. One way to kick off this skill I think is to work with an agent on some task and then after it has finished satisfactorily say something like “ turn this into a skill”
  • Package Managers Need to Cool Down: It looks like good practice for installing dependencies is going to include only using ones that have been released and in use for a period of days so that they can be vetted by security people.
  • My Approach to Building Large Technical Projects: Figure out where to start by breaking down big ideas into smaller chunks and looking for the small-ish, ultimately demo-able steps. Work on the stuff that makes a project usable by yourself.
  • Thoughtworks agent summit: Stuff to think about:
    • Everybody is trying to figure out what’s happening
    • More accumulated debt
      • Technical, code decisions that make future changes hard
      • Cognitive, shared understanding of how the system works in humans drifts
      • Intent debt, documented goals of the system and details drift
      • Drift gets worse as speed of change goes up
    • Humans as supervisors
    • Need for an exact audit trail with rollback
    • If code is made cheap, what’s the important artifact we produce?
  • Relocating Rigor: When we write less code we don’t build up understanding of the system in quite the same way or at all. And yet we’re still required, at this point anyway, to fix it when it brakes. This contradiction is a problem. Where does the rigor go when agents are writing code now? Where is our new touchpoint with the system during development to create confidence that the thing we built does what we want? We should probably get really good at verification.
  • Agent responsibly: Related to above, this team is thinking about guard rails now and verifiers that can say the system works. Also speaks to a responsibility on the part of humans to understand what they’re shipping. “Are you willing to support this change in an incident?”
  • Write Terraform Tests: TIL terraform supports unit and functional testing.