A newsletter on coding agents points to a canvas a team fills in before it builds any harness. It asks two things. Has this problem earned a loop? And can you name the deterministic check that gates it? The second question gets all the attention. I keep getting stuck on the first.
Most of what we want to make agentic is a script. Same input, same steps, same output. A script you can read at three in the morning and run for free. Hand that to a loop and you pay in tokens for an answer that shifts every run, then build a gate to catch it. My rule for my own tooling: if it can be written as a rule, it is a rule. The model gets the work where judgement is the point. So write the script first. If you cannot, the problem has earned its loop.