Autoresearch: Counties Manukau
Autonomous AI agents explored 887 healthcare configurations for Counties Manukau, modelling 50,000 residents interacting with 186 services over 52 weekly ticks.
All data is synthetic. Findings indicate what the model predicts, not what will happen. Validated across multiple seeds where noted.
Agent Fleet Performance
Seven agents explored different objectives simultaneously. Each proposes configuration changes, runs a simulation, and keeps improvements.
Headline Metrics: Baseline vs Best
How far did each agent move its primary metric from baseline?
Improvement Trajectory
Each dot is an accepted experiment. Agents iteratively improve their primary metric.
Causal Links — Ranked by Evidence Strength
Which configuration changes reliably improve outcomes?
Robust findings
Surprising non-effects
Sensitivity Tests
Three policy interventions tested against baseline (3-seed average). All showed measurable improvement on their target metric.
Key Tensions
Equity vs Maori Equity
The equity explorer achieves near-perfect deprivation parity (0.999) but its configuration scores 0.897 on Maori access. The Maori equity explorer achieves 0.945 on Maori access but only 0.978 on deprivation equity. These are partially misaligned goals requiring dual-target strategies.
Telehealth Convergence
Three independent agents (rural, Maori equity, demand) converged on telehealth expansion as their most effective lever. This is the strongest evidence in the fleet: telehealth is a robust, multi-dimensional improvement that simultaneously helps rural access, Maori equity, and unmet demand.
Autoresearch simulation: Counties Manukau, 50K synthetic residents, 186 services, 887 experiments across 7 AI agents. All findings are model predictions on synthetic data.