Building an Advanced Agentic Harness

A new post about building a smarter AI “harness” — basically the control system around an AI so it can plan, check its work, stay on budget, and avoid going off the rails — should have been a dry engineering explainer. Instead, the real action was in the peanut gallery, where readers turned it into a mini civil war over whether these elaborate AI babysitting systems are the future or just fancy busywork.

The article itself argues that one raw AI prompt isn’t enough if you want something reliable. So the author layers on structure: planning ahead, splitting jobs into parallel tasks, keeping different kinds of memory, checking results, tracking cost, and recording everything so mistakes can be replayed later. The demo task is almost comically simple — comparing cities by population, time zone, and short summaries — but that simplicity is the point: it shows how even a basic job can be broken into many smaller steps.

But the comments? Way spicier. One critic came in swinging, basically calling “skills, harnesses, memory systems, whatever” a recycled idea that sounds smart and ends up “totally useless in practice.” Ouch. Another commenter rejected the author’s planning graph entirely, saying they’d rather just give the AI a live coding sandbox and let it improvise. On the flip side, some readers were openly bookmarking it, while others wanted more reading recommendations because they’re building their own versions right now. The biggest recurring side-eye was simple: where are the receipts? Multiple readers wanted proof that all this extra structure actually helps instead of just making the AI slower, pricier, and weirder. In other words: classic tech thread energy — one person says it’s the future, one says it’s fake productivity, and everyone else asks for benchmarks.