Minimal Seed Harness

Start an automatic harness-evolution loop from a deliberately minimal seed so every component the loop adds must earn its place against measured rollouts, not inherit unattributed credit from hand-tuning.

When bootstrapping an automated harness evolution outer loop, the temptation is to seed it from a reasonably capable starting harness so early iterations aren't wasted on obvious gaps. That temptation should be resisted: a seed harness already fitted to the target benchmark contaminates the attribution of every edit that follows, because there is no way to tell whether a later gain came from the loop's own evolution or from credit the hand-tuned seed was carrying all along.

The alternative is a deliberately minimal seed: a single shell-execution tool, no middleware, no skills, no sub-agents — nothing pre-tuned for the target task or benchmark. "The minimal seed forces every component AHE adds to earn its place against measured rollouts." This is the design-time twin of minimal prompt crafting: start from the smallest functional harness, and let every addition beyond it be justified by an evidence-driven change manifest rather than inherited from an unmeasured starting point.

A minimal seed can still be given a small, protected head start — for example a one-shot bootstrap pass that seeds a handful of reusable skills from framework source and public references before the evolve loop begins — as long as that head start gets no special protection afterward: from the next iteration on, the evolve agent is free to keep, refine, or delete those seeded skills based on observed rollouts, the same as anything it wrote itself. A protected head start that the loop can never touch would reintroduce exactly the attribution contamination the minimal seed exists to avoid.

---

Sources

• Agentic Harness Engineering (Lin, Liu, Pan, et al.), §3.1