Most agent harnesses rot because nobody deletes anything
Most agent harnesses rot because rules pile up and nobody checks which ones still prevent a real mistake.
A 6x gap from the harness alone explains why everyone’s “same model” gets such different results..
People tend to focus on the model when making comparisons, and rarely talk about the harness they’re using.
The “mine past corrections into rules” bit is right but the hard part is pruning, not writing. Rules accumulate forever if nobody checks whether they still fire.
If you can’t point to a specific save, cut it. Most harnesses rot because people only ever add.
My carry blocks already do this for open items. Each one ends in one of three ways: it’s done, it’s dropped with a reason, or it’s promoted to the backlog. Rules need the same exit. A rule without an exit keeps occupying your context window.
how do you define best? What metrics are you using: instruction adherence, task completion, context usage, or something else? A smaller kernel may win on efficiency, but does it do what I care about.
If a shorter file drops the single rule that prevents a bad deploy, it scores better on context usage but loses what I wanted. I want fewer rules, each one justified. Vercel has a course on building a harness.
It really shows how much everything around the model matters. IMO it also quietly debunks the “AI will replace devs” narrative, because that stat shows just how much the person piloting the agent makes a difference.
Figuring out what to discard is part of the piloting. Next up, I want to put the rules and guardrails I’ve gathered from community threads and best practices guides to the same test. For now, I’m adding Marmelab’s write-up to my reading list.