The headline on this Superpowers 6 write-up is a 50% cut in build time and 60% fewer tokens. The part worth stealing is that none of it came from a smarter model.

Look at where the savings actually landed. Two review subagents — one checking code quality, one checking spec compliance — got merged into a single reviewer. The reviewers stopped shelling out to git mid-review because the harness now pre-bakes a formatted diff packet up front. A tighter “reviewer contract” cut reviewer output by 41% without changing the verdicts, and a leaner narration recipe shaved another 54% off cost. The orchestrator learned to route easy subtasks to a cheaper model instead of sending everything to the big one. Every one of those is a plumbing change to the verify-and-route layer, not the generation step.

This matches what I see running agentic systems in production: the generation call is rarely the cost center. The review, critique, and re-verification subagents are — and their token spend hides inside verbose contracts nobody budgeted. If you can’t see per-subagent cost in your eval harness, you’re optimizing the wrong call.

The other detail I liked: the winning configuration wasn’t hand-tuned. The author pointed an automated research loop at it and ran 25 pre-registered experiments overnight for about $165, then kept the config that measured best. That’s an agent tuning an agent against a real metric — which is roughly where LLM ops is heading whether we build the harness for it or not.

So here’s the question I keep coming back to: do you actually know which subagent in your pipeline burns the most tokens per successful task? The HN thread has people arguing about model choice — I’d bet most of them can’t answer that one.