Designing for Agentic Implementers: A Capability-Calibrated Methodology for Mixed-Intelligence Software Systems

Software-design methodologies assume a human implementer who closes ambiguities by judgement. When the implementer is a large language model, the detail a specification must carry becomes a function of the implementer's capability. We present AMD (Agentic Mixed-intelligence Design), a methodology in which specification detail is set per concern from an Implementer Profile, and Φ, an executable conformance measure that combines a static check, tests against a counter-example corpus Σ, and an LLM judge bounded by the rule Φ ≤ min(M1, M2). We calibrate the capability-by-detail surface on eight problem classes, with implementers from local open-weight models to frontier API tiers. Four results follow. The test modality carries the discriminating signal at skeleton stage, where the judge saturates and bounds Φ at 0.794. A minimal specification yields trivially passable tests and a spurious pass, a degeneracy any study of specification detail must exclude. An offline judge proxy agrees with a frontier judge within 0.05 for strong API models and breaks down for local models near the capability frontier. The monotone hypothesis, that a more capable implementer needs less detail, holds on one problem class and is suggestive on a second (tie-corrected Spearman ρ = −0.866), saturates within a model family, and fails where the tier design does not straddle the pass threshold (pooled sign test p = 1/16, 1/4 excluding the contaminated class). Specifying the toolkit itself in AMD closed five grammar gaps (Φ = 0.804).