Strategy-first synthesis planning for complex natural products

Retrosynthesis models trained on patent reactions score near-perfectly on benchmarks drawn from the same corpora, yet fail on complex natural products. We present SynthEx, an agentic planner that first commits to a high-level synthetic strategy and then writes each disconnection directly as an atom-level molecular graph edit. A critic--editor loop repairs the resulting route in place while leaving the strategy intact. On 1,098 complex natural products, a near-exhaustively resourced template planner solves 18.4 % of targets, against 63.9 % for SynthEx. The chemistry SynthEx proposes is measurably distinct: it forms a ring in 16.0 % of steps against 2.8 % for a state-of-the-art corpus-trained single-step model, which recovers fewer than a third of SynthEx's disconnections in its top-5. Ten synthetic chemists blind rating reaction steps found no significant difference from published human work on three of four axes. All routes are released as SynthAtlas, an open repository of dated predictions on complex natural products.