How agents actually work

part one — eight problems, one loop

hover a card to see where it attaches

{{ n.num }}

{{ n.label }}

{{ c.num }}

{{ c.src }}

    part two — around the loop, not in it:
    skills·MCP·memory·subagents·prompt injection
  

an agent's output is the change it made to the world

Once you take that seriously, the transcript stops being the thing you evaluate and becomes a debugging artifact. Refusal becomes a positive outcome a test can assert on. A crash mid-run becomes a question about the filesystem.